check_requirements(exclude=('tensorboard', 'pycocotools', 'thop'))
modal_stage_model = None
if opt.modal_stage_model is not None and opt.modal_stage_model != "":
modal_stage_model = None
if opt.modal_stage_model != "":
device = select_device(opt.device, batch_size=opt.batch_size)
modal_ckpt = torch.load(opt.modal_stage_model, map_location=device)
nc = modal_ckpt['model'].yaml['nc']
modal_stage_model = Model(modal_ckpt['model'].yaml, ch=3,
nc=nc).to(device)
state_dict = modal_ckpt['model'].float().state_dict()
state_dict = intersect_dicts(state_dict,
modal_stage_model.state_dict(),
exclude=[])
modal_stage_model.load_state_dict(state_dict, strict=False)
modal_stage_model.eval()
if opt.task in ('train', 'val', 'test'): # run normally
test(opt.data,
opt.weights,
opt.batch_size,
opt.img_size,
opt.conf_thres,
opt.iou_thres,
opt.save_json,
opt.single_cls,
opt.augment,
opt.verbose,
save_txt=opt.save_txt | opt.save_hybrid,
save_hybrid=opt.save_hybrid,
def train(hyp):
print(f'Hyperparameters {hyp}')
log_dir = tb_writer.log_dir if tb_writer else 'runs/evolution' # run directory
wdir = str(Path(log_dir) / 'weights') + os.sep # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir + 'last.pt'
best = wdir + 'best.pt'
results_file = log_dir + os.sep + 'results.txt'
# Save run settings
with open(Path(log_dir) / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(Path(log_dir) / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
epochs = opt.epochs # 300
batch_size = opt.batch_size # 64
weights = opt.weights # initial training weights
# Configure 载入图片路径
init_seeds(1)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg, nc=nc).to(device)
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
# 确定训练和测试图片的尺寸
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if hyp['optimizer'] == 'adam': # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# 定义学习率的变化趋势，这里采用consine learning rate decay.
lf = lambda x: ((
(1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.9 + 0.1 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs, save_dir=log_dir)
# Load Model
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if model.state_dict()[k].shape == v.shape
} # to FP32, filter
model.load_state_dict(ckpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \
"Please delete or update %s and try again, or use --weights '' to train from scratch." \
% (opt.weights, opt.cfg, opt.weights, opt.weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# epochs
start_epoch = ckpt['epoch'] + 1
if epochs < start_epoch:
print(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (opt.weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O1',
verbosity=0)
# Distributed training
if device.type != 'cpu' and torch.cuda.device_count(
) > 1 and dist.is_available():
dist.init_process_group(
backend='nccl', # distributed backend
init_method='tcp://127.0.0.1:9999', # init method
world_size=1, # number of nodes
rank=0) # node rank
# model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) # requires world_size > 1
model = torch.nn.parallel.DistributedDataParallel(model)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (
mlc, nc, opt.cfg)
# Testloader
testloader = create_dataloader(test_path,
imgsz_test,
batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images,
rect=True)[0]
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# Class frequency
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1.
# model._initialize_biases(cf.to(device))
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Check anchors
if not opt.noautoanchor:
check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
# Exponential moving average
ema = torch_utils.ModelEMA(model)
# Start training
t0 = time.time()
nw = max(3 * nb,
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
scheduler.last_epoch = start_epoch - 1 # do not move
print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
print('Using %g dataloader workers' % dataloader.num_workers)
print('Starting training for %g epochs...' % epochs)
# torch.autograd.set_detect_anomaly(True)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 -
maps)**2 # class weights
image_weights = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=w)
dataset.indices = random.choices(range(dataset.n),
weights=image_weights,
k=dataset.n) # rand weighted idx
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls',
'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(
ni, xi,
[0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi,
[0.9, hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets.to(device), model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_cached() /
1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1),
mem, *mloss, targets.shape[0],
imgs.shape[-1])
pbar.set_description(s)
# Plot
if ni < 3:
f = str(Path(log_dir) / ('train_batch%g.jpg' % ni)) # filename
result = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer and result is not None:
tb_writer.add_image(f,
result,
dataformats='HWC',
global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
scheduler.step()
# mAP
ema.update_attr(model,
include=['md', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=batch_size,
imgsz=imgsz_test,
save_json=final_epoch
and opt.data.endswith(os.sep + 'coco.yaml'),
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = [
'train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss',
'val/cls_loss'
]
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
# Strip optimizers
n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.name
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket and ispt else None # upload
# Finish
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1,
(time.time() - t0) / 3600))
dist.destroy_process_group(
) if device.type != 'cpu' and torch.cuda.device_count() > 1 else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None):
logger.info(f'Hyperparameters {hyp}')
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
wdir = log_dir / 'weights' # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = str(log_dir / 'results.txt')
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Save run settings
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
freeze = [
'',
] # parameter names to freeze (full or partial)
if any(freeze):
for k, v in model.named_parameters():
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
v.requires_grad = True
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[
'lrf']) + hyp['lrf'] # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
shutil.copytree(wdir, wdir.parent /
f'weights_backup_epoch{start_epoch - 1}'
) # save previous weights
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# Exponential moving average
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers)[0] # testloader
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
logger.info(
'Image sizes %g train, %g test\nUsing %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs))
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if ni < 3:
f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename
result = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer and result is not None:
tb_writer.add_image(f,
result,
dataformats='HWC',
global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
if final_epoch: # replot predictions
[
os.remove(x) for x in glob.glob(
str(log_dir / 'test_batch*_pred.jpg'))
if os.path.exists(x)
]
results, maps, times = test.test(opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = [
'train/giou_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/giou_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = opt.name if opt.name.isnumeric() else ''
fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt'
for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
if str(f2).endswith('.pt'): # is *.pt
strip_optimizer(f2) # strip optimizer
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket else None # upload
# Finish
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
return results
def detect():
weights, imgsz, output = opt.checkpoint, opt.img_size, opt.output_video
# Initialize
device = torch_utils.select_device(opt.device)
half = device.type != 'cpu' # half precision only supported on CUDA
with open(opt.project) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader)
names = data_dict['names']
dest_object = data_dict[
'dest_object'] if 'dest_object' in data_dict else names
if 'name_map' in data_dict:
name_map = data_dict['name_map']
else:
name_map = defaultdict()
for x in dest_object:
name_map[x] = x
draw_names = set()
for k, v in name_map.items():
draw_names.add(v)
draw_names = list(draw_names)
colors = get_all_colors(len(draw_names))
model = Model(data_dict).to(device)
model.load_state_dict(torch.load(weights, map_location=device))
model.to(device).eval()
if half:
model.half()
if opt.video_type == 'camera':
rtsp = 'rtsp://*****:*****@' + opt.input_video + ':554/h264/chCH/sub/av_stream'
cap = cv2.VideoCapture(rtsp)
elif opt.video_type == 'video':
cap = cv2.VideoCapture(opt.input_video)
else:
print('Input video type ERROR!')
return
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
print(fps, width, height)
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
outstream = cv2.VideoWriter(output, fourcc, fps, (width, height))
t0 = time.time()
img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img
_ = model(img.half() if half else img
) if device.type != 'cpu' else None # run once
while (cap.isOpened()):
ret, im0s = cap.read()
if ret:
pred = forward_one(model, im0s, imgsz, device, half, opt)
for i, det in enumerate(pred): # detections per image
if det is not None and len(det):
for *xyxy, conf, cls in det:
if names[int(cls)] in dest_object:
draw_str = name_map[names[int(cls)]]
color = colors[draw_names.index(draw_str)]
label = '%s %.2f' % (draw_str, conf)
xmin, ymin, xmax, ymax = xyxy
cv2.rectangle(im0s, (xmin, ymin), (xmax, ymax),
color=color,
thickness=1)
cv2.putText(im0s, label, (xmin, ymax),
cv2.FONT_HERSHEY_SIMPLEX, 0.6, color,
1)
if opt.need_view:
cv2.imshow('real-time', im0s)
key = cv2.waitKey(delay=1)
if key == ord('q'):
break
else:
outstream.write(im0s)
cv2.destroyAllWindows()
print('Done. (%.3fs)' % (time.time() - t0))
def train(hyp, opt, device, tb_writer=None):
logger.info(f'Hyperparameters {hyp}')
# 获取记录训练日志的路径
# 如果设置进化算法则不会传入tb_writer(则为None)，设置一个evolve文件夹作为日志目录
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
# 设置保存权重的路径
wdir = log_dir / 'weights' # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir / 'last.pt'
best = wdir / 'best.pt'
# 设置保存results的路径
results_file = str(log_dir / 'results.txt')
# 获取轮次、批次、总批次(涉及到分布式训练)、权重、进程序号(主要用于分布式训练)
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# rank = -1
# Save run settings
# 保存hyp和opt
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
cuda = (device.type != 'cpu')
# 设置随机种子
init_seeds(2 + rank)
with open(opt.data) as f: # 加载数据配置信息
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(
rank): # torch_distributed_zero_first同步所有进程
check_dataset(
data_dict
) # check_dataset检查数据集，如果没找到数据集则下载数据集(仅适用于项目中自带的yaml文件数据集)
# 获取训练集、测试集图片路径
train_path = data_dict['train']
test_path = data_dict['val']
# 获取类别数量和类别名字, 如果设置了opt.single_cls则为一类
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained: # 如果采用预训练
# 加载模型，从google云盘中自动下载模型
# 但通常会下载失败，建议提前下载下来放进weights目录
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
# 加载检查点
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
"""
这里模型创建，可通过opt.cfg，也可通过ckpt['model'].yaml
这里的区别在于是否是resume，resume时会将opt.cfg设为空，则按照ckpt['model'].yaml创建模型；
这也影响着下面是否除去anchor的key(也就是不加载anchor)，如果resume则不加载anchor
主要是因为保存的模型会保存anchors，有时候用户自定义了anchor之后，再resume，则原来基于coco数据集的anchor就会覆盖自己设定的anchor，
参考https://github.com/ultralytics/yolov5/issues/459
所以下面设置了intersect_dicts，该函数就是忽略掉exclude
"""
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
# 显示加载预训练权重的的键值对和创建模型的键值对
# 如果设置了resume，则会少加载两个键值对(anchors,anchor_grid)
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
# 创建模型， ch为输入图片通道
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
"""
冻结模型层,设置冻结层名字即可
具体可以查看https://github.com/ultralytics/yolov5/issues/679
但作者不鼓励冻结层,因为他的实验当中显示冻结层不能获得更好的性能,参照:https://github.com/ultralytics/yolov5/pull/707
并且作者为了使得优化参数分组可以正常进行,在下面将所有参数的requires_grad设为了True
其实这里只是给一个freeze的示例
"""
freeze = [
'',
] # parameter names to freeze (full or partial)
if any(freeze):
for k, v in model.named_parameters():
# print(k,v)
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
"""
nbs为模拟的batch_size;
就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64，
也就是模型梯度累积了64/16=4(accumulate)次之后
再更新一次模型，变相的扩大了batch_size
"""
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing accumulate = 4
# 根据accumulate设置权重衰减系数
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
# 将模型分成三组(weight、bn, bias, 其他所有参数)优化
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
# print(k)
v.requires_grad = True
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
# 选用优化器，并设置pg0组的优化方式
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
# 设置weight、bn的优化方式
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
# 设置biases的优化方式
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
# 打印优化信息
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# 设置学习率衰减，这里为余弦退火方式进行衰减
# 就是根据以下公式lf,epoch和超参数hyp['lrf']进行衰减
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[
'lrf']) + hyp['lrf'] # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Resume
# 初始化开始训练的epoch和最好的结果
# best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, [email protected], [email protected]:0.95]再求和所得
# 根据best_fitness来保存best.pt
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
# 加载优化器与 best_fitness
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
# 加载训练结果result.txt
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
# 加载训练的轮次
# print(ckpt['epoch'])
start_epoch = ckpt['epoch'] + 1 # ckpt['epoch'] = -1
"""
如果resume,则备份权重
尽管目前resume能够近似100%成功的起作用了,参照:https://github.com/ultralytics/yolov5/pull/756
但为了防止resume时出现其他问题,把之前的权重覆盖了,所以这里进行备份,参照:https://github.com/ultralytics/yolov5/pull/765
"""
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
shutil.copytree(wdir, wdir.parent /
f'weights_backup_epoch{start_epoch - 1}'
) # save previous weights
"""
如果新设置epochs小于加载的epoch，
则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数
"""
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
# 获取模型总步长和模型输入图片分辨率
gs = int(max(model.stride)) # grid size (max stride)
# 检查输入图片分辨率确保能够整除总步长gs
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# imgsz, imgsz_test 都是640
# DP mode
# 分布式训练,参照:https://github.com/ultralytics/yolov5/issues/475
# DataParallel模式,仅支持单机多卡
# rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式
# rank=-1且gpu数量=1时,不会进行分布式
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model) # 执行了
# SyncBatchNorm
# 使用跨卡同步BN
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# Exponential moving average 指数滑动平均，或指数加权平均
# 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
# 如果rank不等于-1,则使用DistributedDataParallel模式
# local_rank为gpu编号,rank为进程,例如rank=3，local_rank=0 表示第 3 个进程内的第 1 块 GPU。
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
# 创建训练集dataloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers)
"""
获取标签中最大的类别值，并于类别数作比较
如果小于类别数则表示有问题
"""
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
# 更新ema模型的updates参数,保持ema的平滑性
ema.updates = start_epoch * nb // accumulate # set EMA updates
# 创建测试集dataloader
testloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers)[0] # testloader
if not opt.resume:
# 将所有样本的标签拼接到一起shape为(total, 5)，统计后做可视化
labels = np.concatenate(dataset.labels, 0)
# 获得所有样本的类别
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
# 根据上面的统计对所有样本的类别，中心点xy位置，长宽wh做可视化
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Anchors
"""
计算默认锚点anchor与数据集标签框的长宽比值
标签的长h宽w与anchor的长h_a宽w_a的比值, 即h/h_a, w/w_a都要在(1/hyp['anchor_t'], hyp['anchor_t'])是可以接受的
如果标签框满足上面条件的数量小于总数的99%，则根据k-mean算法聚类新的锚点anchor
"""
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
# 根据自己数据集的类别数设置分类损失的系数
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
# 设置类别数，超参数
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
"""
设置giou的值在objectness loss中做标签的系数, 使用代码如下
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)
这里model.gr=1，也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签
"""
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
# 根据labels初始化图片采样权重
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
# 获取类别的名字
model.names = names
# Start training
t0 = time.time()
# 获取热身训练的迭代次数
nw = max(round(hyp['warmup_epochs'] * nb),
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
# 初始化mAP和results
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
"""
设置学习率衰减所进行到的轮次，
目的是打断训练后，--resume接着训练也能正常的衔接之前的训练进行学习率衰减
"""
scheduler.last_epoch = start_epoch - 1 # do not move
# 通过torch1.6自带的api设置混合精度训练
scaler = amp.GradScaler(enabled=cuda)
"""
打印训练和测试输入图片分辨率
加载图片时调用的cpu进程数
从哪个epoch开始训练
"""
logger.info(
'Image sizes %g train, %g test\nUsing %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs))
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
"""
如果设置进行图片采样策略，
则根据前面初始化的图片采样权重model.class_weights以及maps配合每张图片包含的类别数
通过random.choices生成图片索引indices从而进行采样
"""
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
# 如果是DDP模式,则广播采样策略
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
# 广播索引到其他group
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
# 初始化训练时打印的平均损失信息
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
# DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子，
# 每次epoch不同，随机种子就不同
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
# tqdm 创建进度条，方便训练时 信息的展示
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
# 计算迭代的次数iteration
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
"""
热身训练(前nw次迭代)
在前nw次迭代中，根据以下方式选取accumulate和学习率
"""
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
"""
bias的学习率从0.1下降到基准学习率lr*lf(epoch)，
其他的参数学习率从0增加到lr*lf(epoch).
lf为上面设置的余弦退火的衰减函数
"""
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
# 动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937)
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
# 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
# 混合精度
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward 前向传播
# Loss
# 计算损失，包括分类损失，objectness损失，框的回归损失
# loss为总损失值，loss_items为一个元组，包含分类损失，objectness损失，框的回归损失和总损失
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
if rank != -1:
# 平均不同gpu之间的梯度
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
# 反向传播
scaler.scale(loss).backward()
# Optimize
# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
# 打印显存，进行的轮次，损失，target的数量和图片的size等信息
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
# 进度条显示以上信息
pbar.set_description(s)
# Plot
# 将前三次迭代batch的标签框在图片上画出来并保存
if ni < 3:
f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename
result = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer and result is not None:
tb_writer.add_image(f,
result,
dataformats='HWC',
global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
# 进行学习率衰减
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
# 更新EMA的属性
# 添加include的属性
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
# 判断该epoch是否为最后一轮
final_epoch = epoch + 1 == epochs
# 对测试集进行测试，计算mAP等指标
# 测试时使用的是EMA模型
if not opt.notest or final_epoch: # Calculate mAP
if final_epoch: # replot predictions
[
os.remove(x) for x in glob.glob(
str(log_dir / 'test_batch*_pred.jpg'))
if os.path.exists(x)
]
results, maps, times = test.test(opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir)
# Write
# 将指标写入result.txt
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
# 如果设置opt.bucket, 上传results.txt到谷歌云盘
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
# 添加指标，损失等信息到tensorboard显示
if tb_writer:
tags = [
'train/giou_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/giou_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
# 更新best_fitness
fi = fitness(np.array(results).reshape(
1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
"""
保存模型，还保存了epoch，results，optimizer等信息，
optimizer将不会在最后一轮完成后保存
model保存的是EMA的模型
"""
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
"""
模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；
并且对模型进行model.half(), 将Float32的模型->Float16，
可以减少模型大小，提高inference速度
"""
n = opt.name if opt.name.isnumeric() else ''
fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt'
for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
if str(f2).endswith('.pt'): # is *.pt
strip_optimizer(f2) # strip optimizer
# 上传结果到谷歌云盘
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket else None # upload
# Finish
# 可视化results.txt文件
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
# 释放显存
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None, wandb=None):
# lr setting
# hyp['lr0'] = 0.01
min_lr = hyp['lr0'] * hyp['lrf']
logger.info(f'Hyperparameters {hyp}')
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
# train_path = data_dict['train']
# test_path = data_dict['val']
train_path = [
data_dict['train'], data_dict['wider_person_train'],
data_dict['crowd_person_train'], data_dict['local1111_train'],
data_dict['background'], data_dict['exdark_train']
]
#train_path = [data_dict['part_train'], data_dict['part_wider_person_train'], data_dict['part_crowd_person_train'],
# data_dict['part_local1111_train'], data_dict['background'], data_dict['part_exdark_train']]
test_path = [
data_dict['val'], data_dict['wider_person_val'],
data_dict['crowd_person_val'], data_dict['local1111_val'],
data_dict['exdark_val']
]
# for test
train_path = [data_dict['background'], data_dict['part_exdark_train']]
test_path = [data_dict['local1111_val']]
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
# freeze = [] # parameter names to freeze (full or partial)
freeze = ['model.%s.' % x for x in range(10)
] # parameter names to freeze (full or partial) #backbone
freeze2 = ['model.%s.' % x for x in range(3)] # fronze first stage 0-3
# freeze = ['model.%s.' % x for x in range(24)] # parameter names to freeze (full or partial)
# for k, v in model.named_parameters(): # model.0.conv.conv.weight model.0.conv.bn.weight model.0.conv.bn.bias
# print(k)
# exit(1)
# print("*"*20)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze) and "bn" in k: # freeze fbn layer
print('freezing bn %s' % k)
v.requires_grad = False
if any(x in k for x in freeze2):
print('freezing first stage %s' % k)
v.requires_grad = False
exit(1)
# exit(1)
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
# lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine
# scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project='YOLOv5'
if opt.project == 'runs/train' else Path(opt.project).stem,
name=save_dir.stem,
id=ckpt.get('wandb_id') if 'ckpt' in locals() else None,
mode="offline")
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
# if ckpt['optimizer'] is not None:
# optimizer.load_state_dict(ckpt['optimizer'])
# best_fitness = ckpt['best_fitness']
# Results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
# start_epoch = ckpt['epoch'] + 1
start_epoch = 0
# print("resume start_epoch:", start_epoch)
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers)[0] # testloader
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
plot_labels(labels, save_dir=save_dir)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
if wandb:
wandb.log({
"Labels": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('*labels*.png')
]
})
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
# scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
logger.info('Image sizes %g train, %g test\n'
'Using %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs))
# add CosineAnnealingLR for swa
cut_b = int(len(dataloader) /
accumulate) * accumulate # only process [0:cut_b] data
t_max = len(dataloader) // accumulate - 1 # lr period
scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
T_max=t_max,
eta_min=min_lr)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
lr_list = []
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
scheduler.last_epoch = -1
scheduler.step()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
# cut data loader
if i >= cut_b:
continue
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# # Warmup
# if ni <= nw:
# xi = [0, nw] # x interp
# # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
# accumulate = max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round())
# for j, x in enumerate(optimizer.param_groups):
# # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
# x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
# if 'momentum' in x:
# x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# Optimize
if (i + 1) % accumulate == 0:
# mark down lr for tensor board
lr = [x['lr'] for x in optimizer.param_groups]
lr_list.append(lr[-1])
tags = ['x/lr0', 'x/lr1', 'x/lr2'] # params
for x, tag in zip(lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, ni)
if wandb:
wandb.log({tag: x}) # W&B
# optimizer update
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# update lr
scheduler.step()
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if plots and ni < 3:
f = save_dir / f'train_batch{ni}.jpg' # filename
plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif plots and ni == 3 and wandb:
wandb.log({
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg')
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# # Scheduler
# lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
# scheduler.step()
# check lr
assert lr_list[-1] == min_lr # make sure the last lr is min_lr
assert lr_list[0] == hyp['lr0'] # make sure the first lr is max_lr
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=model, #ema.ema
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch,
log_imgs=opt.log_imgs if wandb else 0)
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
# ckpt = {'epoch': epoch,
# 'best_fitness': best_fitness,
# 'training_results': f.read(),
# 'model': ema.ema,
# 'optimizer': None if final_epoch else optimizer.state_dict(),
# 'wandb_id': wandb_run.id if wandb else None}
ckpt_origin = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
model,
'optimizer':
None if final_epoch else optimizer.state_dict(),
'wandb_id':
wandb_run.id if wandb else None
}
# Save last, best and delete
# torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt_origin, best)
# save each epoch model
torch.save(
ckpt_origin,
os.path.join(
os.path.split(last)[0], "swa_" + str(epoch) + ".pt"))
# del ckpt
del ckpt_origin
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = opt.name if opt.name.isnumeric() else ''
fresults, flast, fbest = save_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt'
for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file],
[flast, fbest, fresults]):
if f1.exists():
os.rename(f1, f2) # rename
if str(f2).endswith('.pt'): # is *.pt
strip_optimizer(f2) # strip optimizer
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket else None # upload
# Finish
if plots:
plot_results(save_dir=save_dir) # save as results.png
if wandb:
wandb.log({
"Results": [
wandb.Image(str(save_dir / x), caption=x)
for x in ['results.png', 'precision-recall_curve.png']
]
})
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
else:
dist.destroy_process_group()
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None, wandb=None):
logger.info(f'Hyperparameters {hyp}')
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings,超参数，训练para
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(
rank): # torch_distributed_zero_first同步所有进程
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
# 所以这里主要是设定一个，如果加载预训练权重进行训练的话，就去除掉权重中的anchor，采用用户自定义的；
# 如果是resume的话，就是不去除anchor，就权重和anchor一起加载， 接着训练；
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create resume时将opt.cfg设为空
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys 如果resume，则加载权重中保存的anchor来继续训练；
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info('Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights))
# 显示加载预训练权重的的键值对和创建模型的键值对
# 如果设置了resume，则会少加载两个键值对(anchors,anchor_grid)
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
#将模型分成三组(weight、bn, bias, 其他所有参数)优化
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[
'lrf']) + hyp['lrf'] # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project='YOLOv5'
if opt.project == 'runs/train' else Path(opt.project).stem,
name=save_dir.stem,
id=ckpt.get('wandb_id') if 'ckpt' in locals() else None)
loggers = {'wandb': wandb} # loggers dict
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode DataParallel模式,仅支持单机多卡
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# Exponential moving average 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
# 如果rank不等于-1,则使用DistributedDataParallel模式
# local_rank为gpu编号,rank为进程,例如rank=3，local_rank=0 表示第 3 个进程内的第 1 块 GPU。
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
"""
获取标签中最大的类别值，并于类别数作比较
如果小于类别数则表示有问题
"""
# Process 0
if rank in [-1, 0]:
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5)[0]
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
# 根据上面的统计对所有样本的类别，中心点xy位置，长宽wh做可视化
if plots:
Thread(target=plot_labels,
args=(labels, save_dir, loggers),
daemon=True).start()
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda) # 通过torch1.6自带的api设置混合精度训练
logger.info('Image sizes %g train, %g test\n'
'Using %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs))
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0) # 广播索引到其他group
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
# DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子，
# 每次epoch不同，随机种子就不同
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
"""
bias的学习率从0.1下降到基准学习率lr*lf(epoch)，
其他的参数学习率从0增加到lr*lf(epoch).
lf为上面设置的余弦退火的衰减函数
"""
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# loss为总损失值，loss_items为一个元组，包含分类损失，objectness损失，框的回归损失和总损失
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if plots and ni < 3:
f = save_dir / f'train_batch{ni}.jpg' # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif plots and ni == 3 and wandb:
wandb.log({
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg')
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch,
log_imgs=opt.log_imgs if wandb else 0)
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict(),
'wandb_id':
wandb_run.id if wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
# 模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；
# 并且对模型进行model.half(), 将Float32的模型->Float16，
if rank in [-1, 0]:
# Strip optimizers
for f in [last, best]:
if f.exists(): # is *.pt
strip_optimizer(f) # strip optimizer
os.system('gsutil cp %s gs://%s/weights' %
(f, opt.bucket)) if opt.bucket else None # upload
# Plots
if plots:
plot_results(save_dir=save_dir) # save as results.png
if wandb:
files = [
'results.png', 'precision_recall_curve.png',
'confusion_matrix.png'
]
wandb.log({
"Results": [
wandb.Image(str(save_dir / f), caption=f)
for f in files if (save_dir / f).exists()
]
})
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
# Test best.pt
if opt.data.endswith('coco.yaml') and nc == 80: # if COCO
results, _, _ = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=attempt_load(best if best.exists() else last,
device).half(),
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
save_json=True, # use pycocotools
plots=False)
else:
dist.destroy_process_group()
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def detect(save_img=False):
out, source, weights, view_img, save_txt, imgsz = \
opt.output, opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size
webcam = source.isnumeric() or source.startswith(
'rtsp') or source.startswith('http') or source.endswith('.txt')
# Initialize
set_logging()
device = select_device(opt.device)
if os.path.exists(out):
shutil.rmtree(out) # delete output folder
os.makedirs(out) # make new output folder
half = device.type != 'cpu' # half precision only supported on CUDA
# Load model
# add by zlf at 20201027
# load FP16 model
# model=torch.load(weights)['model']
# for n,p in model.named_parameters():
# print(p.dtype)
# model = attempt_load(weights, map_location=device) # load FP32 model
# imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size
# 20201019 load model by zlf
# new method of loading weight;only for 'torch.save(model,state_dict())'
net = Model('./models/yolov5s.yaml').fuse()
# net = ModelEMA(net)
# net = net.ema.cuda()
state_dict = torch.load('20201110.pth', map_location=torch.device('cpu'))
# state_dict = torch.jit.load('n.pt', map_location=torch.device('cpu'))
# model = state_dict
# model.cuda()
model_dict = net.state_dict()
for k, v in state_dict.items():
name = k[0:] # remove `module.`
model_dict[name] = v
net.load_state_dict(model_dict, strict=True)
model = net.half().cuda()
# model.eval()
# input1 = torch.rand((1,3,192,320)).cuda()
model.eval()
# torch.jit.trace(model,input1).save('new2.pt')
imgsz = 320
# add by zlf at 20201009
if half:
model.half() # to FP16
# Second-stage classifier
classify = False
if classify:
modelc = load_classifier(name='resnet101', n=2) # initialize
modelc.load_state_dict(
torch.load('weights/resnet101.pt',
map_location=device)['model']) # load weights
modelc.to(device).eval()
# Set Dataloader
vid_path, vid_writer = None, None
if webcam:
view_img = True
cudnn.benchmark = True # set True to speed up constant image size inference
dataset = LoadStreams(source, img_size=imgsz)
else:
save_img = True
#TODO:cudnn.benchmark = True
dataset = LoadImages(source, img_size=imgsz)
# Get names and colors
# names = model.module.names if hasattr(model, 'module') else model.names
names = [
'CAR', 'CARPLATE', 'BICYCLE', 'TRICYCLE', 'PEOPLE', 'MOTORCYCLE',
'LOGO_AUDI', 'LOGO_BENZE', 'LOGO_BENZC', 'LOGO_BMW', 'LOGO_BUICK',
'LOGO_CHEVROLET', 'LOGO_CITROEN', 'LOGO_FORD', 'LOGO_HONDA',
'LOGO_HYUNDAI', 'LOGO_KIA', 'LOGO_MAZDA', 'LOGO_NISSAN',
'LOGO_PEUGEOT', 'LOGO_SKODA', 'LOGO_SUZUKI', 'LOGO_TOYOTA',
'LOGO_VOLVO', 'LOGO_VW', 'LOGO_ZHONGHUA', 'LOGO_SUBARU', 'LOGO_LEXUS',
'LOGO_CADILLAC', 'LOGO_LANDROVER', 'LOGO_JEEP', 'LOGO_BYD',
'LOGO_BYDYUAN', 'LOGO_BYDTANG', 'LOGO_CHERY', 'LOGO_CARRY',
'LOGO_HAVAL', 'LOGO_GREATWALL', 'LOGO_GREATWALLOLD', 'LOGO_ROEWE',
'LOGO_JAC', 'LOGO_HAFEI', 'LOGO_SGMW', 'LOGO_CASY', 'LOGO_CHANAJNX',
'LOGO_CHANGAN', 'LOGO_CHANA', 'LOGO_CHANGANCS', 'LOGO_XIALI',
'LOGO_FAW', 'LOGO_YQBT', 'LOGO_REDFLAG', 'LOGO_GEELY', 'LOGO_EMGRAND',
'LOGO_GLEAGLE', 'LOGO_ENGLON', 'LOGO_BAOJUN', 'LOGO_DF', 'LOGO_JINBEI',
'LOGO_BAIC', 'LOGO_WEIWANG', 'LOGO_HUANSU', 'LOGO_FOTON', 'LOGO_HAIMA',
'LOGO_ZOTYEAUTO', 'LOGO_MITSUBISHI', 'LOGO_RENAULT', 'LOGO_MG',
'LOGO_DODGE', 'LOGO_FIAT', 'LOGO_INFINITI', 'LOGO_MINI', 'LOGO_TESLA',
'LOGO_SMART', 'LOGO_BORGWARD', 'LOGO_JAGUAR', 'LOGO_HUMMER',
'LOGO_PORSCHE', 'LOGO_LAMBORGHINI', 'LOGO_DS', 'LOGO_CROWN',
'LOGO_LUXGEN', 'LOGO_ACURA', 'LOGO_LINCOLN', 'LOGO_SOUEAST',
'LOGO_VENUCIA', 'LOGO_TRUMPCHI', 'LOGO_LEOPAARD', 'LOGO_ZXAUTO',
'LOGO_LIFAN', 'LOGO_HUANGHAI', 'LOGO_HAWTAI', 'LOGO_REIZ',
'LOGO_CHANGHE', 'LOGO_GOLDENDRAGON', 'LOGO_YUTONG', 'LOGO_HUIZHONG',
'LOGO_JMC', 'LOGO_JMCYUSHENG', 'LOGO_LANDWIND', 'LOGO_NAVECO',
'LOGO_QOROS', 'LOGO_OPEL', 'LOGO_YUEJING'
]
colors = [[random.randint(0, 255) for _ in range(3)]
for _ in range(len(names))]
# Run inference
t0 = time.time()
img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img
# _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once
for path, img, im0s, vid_cap in dataset:
f1.write('%s:' % (path.split('/')[-1]))
img = torch.from_numpy(img).to(device)
img = img.half() if half else img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
if img.ndimension() == 3:
img = img.unsqueeze(0)
# Inference
t1 = time_synchronized()
pred = model(img, augment=opt.augment)[0]
# pred = model(img.cuda())
# pred = list(pred)
# pred = detectz(pred)
# Apply NMS
pred = non_max_suppression(pred,
opt.conf_thres,
opt.iou_thres,
classes=opt.classes,
agnostic=opt.agnostic_nms)
t2 = time_synchronized()
# Apply Classifier
if classify:
pred = apply_classifier(pred, modelc, img, im0s)
# Process detections
for i, det in enumerate(pred): # detections per image
if webcam: # batch_size >= 1
p, s, im0 = path[i], '%g: ' % i, im0s[i].copy()
else:
p, s, im0 = path, '', im0s
save_path = str(Path(out) / Path(p).name)
txt_path = str(Path(out) / Path(p).stem) + (
'_%g' % dataset.frame if dataset.mode == 'video' else '')
s += '%gx%g ' % img.shape[2:] # print string
gn = torch.tensor(im0.shape)[[1, 0, 1,
0]] # normalization gain whwh
if det is not None and len(det):
# Rescale boxes from img_size to im0 size
det[:, :4] = scale_coords(img.shape[2:], det[:, :4],
im0.shape).round()
# Print results
for c in det[:, -1].unique():
n = (det[:, -1] == c).sum() # detections per class
s += '%g %ss, ' % (n, names[int(c)]) # add to string
# Write results
for *xyxy, conf, cls, obj_conf, cls_conf in reversed(
det): # add by zlf at 20201026
# add by zlf at 20201019
x1 = int(xyxy[0].item())
y1 = int(xyxy[1].item())
x2 = int(xyxy[2].item())
y2 = int(xyxy[3].item())
f1.write(
"[%s,%.2f,%d,%d,%d,%d]" %
(names[int(cls.item())], round(
(conf.item() * 100), 2), x1, y1, x2, y2))
# add by zlf at 20201019
if save_txt: # Write to file
xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) /
gn).view(-1).tolist() # normalized xywh
with open(txt_path + '.txt', 'a') as f:
f.write(('%g ' * 5 + '\n') %
(cls, *xywh)) # label format
if save_img or view_img: # Add bbox to image
label = '%s|%.2f|%.2f|%.2f' % (names[int(cls)], conf,
obj_conf, cls_conf)
plot_one_box(xyxy,
im0,
label=label,
color=colors[int(cls)],
line_thickness=3)
f1.write('\n')
# Print time (inference + NMS)
print('%sDone. (%.3fs)' % (s, t2 - t1))
# Stream results
if view_img:
cv2.imshow(p, im0)
if cv2.waitKey(1) == ord('q'): # q to quit
raise StopIteration
# Save results (image with detections)
if save_img:
if dataset.mode == 'images':
cv2.imwrite(save_path, im0)
else:
if vid_path != save_path: # new video
vid_path = save_path
if isinstance(vid_writer, cv2.VideoWriter):
vid_writer.release(
) # release previous video writer
fourcc = 'mp4v' # output video codec
fps = vid_cap.get(cv2.CAP_PROP_FPS)
w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
vid_writer = cv2.VideoWriter(
save_path, cv2.VideoWriter_fourcc(*fourcc), fps,
(w, h))
vid_writer.write(im0)
if save_txt or save_img:
print('Results saved to %s' % Path(out))
if platform.system() == 'Darwin' and not opt.update: # MacOS
os.system('open ' + save_path)
print('Done. (%.3fs)' % (time.time() - t0))
def train(hyp, opt, device, tb_writer=None):
print(f'Hyperparameters {hyp}')
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
wdir = str(log_dir / 'weights') + os.sep # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir + 'last.pt'
best = wdir + 'best.pt'
results_file = str(log_dir / 'results.txt')
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# TODO: Use DDP logging. Only the first process is allowed to log.
# Save run settings
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
root_path = data_dict['root']
train_path = data_dict['train']
test_emb_path = data_dict['test_emb']
test_path = data_dict['test']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Remove previous results
if rank in [-1, 0]:
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg, nc=nc).to(device)
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz = imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
# default DDP implementation is slow for accumulation according to: https://pytorch.org/docs/stable/notes/ddp.html
# all-reduce operation is carried out during loss.backward().
# Thus, there would be redundant all-reduce communications in a accumulation procedure,
# which means, the result is still right but the training speed gets slower.
# TODO: If acceleration is needed, there is an implementation of allreduce_post_accumulation
# in https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/BERT/run_pretraining.py
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((
(1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# Load Model
with torch_distributed_zero_first(rank):
attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
exclude = ['anchor'] # exclude keys
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if k in model.state_dict() and not any(x in k for x in exclude)
and model.state_dict()[k].shape == v.shape
}
model.load_state_dict(ckpt['model'], strict=False)
print('Transferred %g/%g items from %s' %
(len(ckpt['model']), len(model.state_dict()), weights))
except KeyError as e:
s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \
"Please delete or update %s and try again, or use --weights '' to train from scratch." \
% (weights, opt.cfg, weights, weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# epochs
start_epoch = ckpt['epoch'] + 1
if epochs < start_epoch:
print(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
print('Using SyncBatchNorm()')
# Exponential moving average
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=(opt.local_rank))
# Trainloader
dataloader, dataset = create_dataloader(root_path,
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
local_rank=rank,
world_size=opt.world_size)
# Testloader
testloader = create_dataloader(root_path,
test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images,
rect=True,
local_rank=-1,
world_size=opt.world_size)[0]
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# model.nID = dataset.nID
# Class frequency
if rank in [-1, 0]:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Check anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Start training
t0 = time.time()
nw = max(3 * nb,
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
if rank in [0, -1]:
print('Image sizes {} train, {} test'.format(str(imgsz),
str(imgsz_test)))
print('Using %g dataloader workers' % dataloader.num_workers)
print('Starting training for %g epochs...' % epochs)
# torch.autograd.set_detect_anomaly(True)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
mloss = torch.zeros(5, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
if rank in [-1, 0]:
print(('\n' + '%10s' * 8 + '%13s') %
('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'lid', 'total',
'targets', 'img_size'))
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(
ni, xi,
[0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi,
[0.9, hyp['momentum']])
# Multi-scale
if opt.multi_scale and random.random() < 0.5:
candidate_shapes = [[608, 1088], [480, 864], [320, 576],
[512, 960], [384, 640]]
curr_shapes = candidate_shapes[random.randint(0, 4)]
imgs = F.interpolate(imgs,
size=curr_shapes,
mode='bilinear',
align_corners=False)
# Autocast
with amp.autocast(enabled=cuda):
# Forward
pred_detect, pred_emb = model(imgs)
# Loss
loss, loss_items = compute_loss(pred_detect, pred_emb,
targets.to(device),
model) # scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema is not None:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.4g' * 6 +
"%10s") % ('%g/%g' % (epoch, epochs - 1), mem, *mloss,
targets.shape[0], ' [%g,%g]' %
(imgs.shape[-1], imgs.shape[-2]))
pbar.set_description(s)
# Plot
if ni < 3:
f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename
result = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer and result is not None:
tb_writer.add_image(f,
result,
dataformats='HWC',
global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema is not None:
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
save_json=final_epoch
and opt.data.endswith(os.sep + 'coco.yaml'),
model=ema.ema.module
if hasattr(ema.ema, 'module') else ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir,
emb_dim=model.module.emb_dim)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 8 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = [
'train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss',
'val/cls_loss'
]
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema.module if hasattr(ema, 'module') else ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = ('_'
if len(opt.name) and not opt.name.isnumeric() else '') + opt.name
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (
f2, opt.bucket)) if opt.bucket and ispt else None # upload
# Finish
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
print('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None, wandb=None):
logger.info(f'Hyperparameters {hyp}')
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
wdir = log_dir / 'weights' # weights dfirectory
os.makedirs(wdir, exist_ok=True)
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = str(log_dir / 'results.txt')
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs_init, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Save run settings
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
if opt.reg_lambda != 0:
# the regularization is based on Synaptic Intelligence as described in the
# paper. ewcData is a list of two elements (best parametes, importance)
# while synData is a dictionary with all the trajectory data needed by SI
model.ewcData, model.synData = create_syn_data(model)
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[
'lrf']) + hyp['lrf'] # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if wandb and wandb.run is None:
id = ckpt.get('wandb_id') if 'ckpt' in locals() else None
wandb_run = wandb.init(config=opt,
resume="allow",
project="YOLOv5",
name=os.path.basename(log_dir),
id=id)
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
shutil.copytree(wdir, wdir.parent /
f'weights_backup_epoch{start_epoch - 1}'
) # save previous weights
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
all_test_dataloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
n_batch=-1)[0]
root = 'G:/projects/core50_350_1f/batches/'
paths = os.listdir(root)
train_paths = []
valid_paths = []
for p in paths:
if 'train' in p:
train_paths.append(root + p)
elif 'val' in p:
valid_paths.append(root + p)
else:
print(p)
# external_memory = ext_memory()
extMem = externalMemory()
for core_batch in range(11):
# Trainloader
if opt.reg_lambda != 0:
init_batch(model, model.ewcData, model.synData)
print(f'------------CORE50 itertaion №:{core_batch}------------')
external_files_path = extMem.file
if core_batch > 0:
train_path = [train_paths[core_batch], external_files_path]
else:
train_path = train_paths[core_batch]
extMem.update_memory(train_paths[core_batch],
update_iters=10 if core_batch == 0 else 1)
dataloader, dataset = create_dataloader(
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
testloader = create_dataloader(valid_paths[core_batch],
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images
and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers)[0] # testloader
#if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.IntTensor(labels[:, 0]) # classes
plot_labels(labels, save_dir=log_dir)
print(torch.bincount(c))
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, core_batch)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(
round(hyp['warmup_epochs'] * nb),
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
logger.info(
'Image sizes %g train, %g test\n'
'Using %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs))
# update number of epochs to iterative training
if core_batch != 0:
epochs = opt.epochs_iter
# x_train, y_train = dataset.get_all_data()
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
mloss = torch.zeros(4, device=device) # mean losses
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'targets', 'img_size'))
# x_train_splitted = torch.split(x_train, 4)
# y_train_splitted = torch.split(y_train, 4)
# pbar = enumerate(zip(x_train_splitted, y_train_splitted))
pbar = enumerate(dataloader)
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, _, _
) in pbar: # batch -------------------------------------------------------------
# imgs = x_train[i * batch_size:(i + 1) * batch_size]
# targets = y_train[i * batch_size:(i + 1) * batch_size]
#
# # preprocess tensor to proper form
# # img, label = zip(imgs, targets) # transposed
# for i, l in enumerate(targets):
# l[:, 0] = i # add target image index for build_targets()
#
# imgs = torch.stack(imgs)
# targets = torch.cat(targets)
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
if opt.reg_lambda != 0:
pre_update(model, model.synData)
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi,
[hyp['warmup_momentum'], hyp['momentum']])
# # Multi-scale
# if opt.multi_scale:
# sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size
# sf = sz / max(imgs.shape[2:]) # scale factor
# if sf != 1:
# ns = [math.ceil(x * sf / gs) * gs for x in
# imgs.shape[2:]] # new shape (stretched to gs-multiple)
# imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
if opt.reg_lambda != 0:
post_update(model, model.synData)
scaler.update()
optimizer.zero_grad()
# if ema:
# ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
# model=ema.ema,
model=model,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir,
plots=epoch == 0, # plot first and last
log_imgs=opt.log_imgs)
# wandb.log({'per class/AP per class': maps})
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/giou_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/giou_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
save = not opt.nosave
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
# 'model': ema.ema,
'model': model,
'optimizer': optimizer.state_dict(),
'wandb_id': wandb_run.id if wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
#consolidate_weights(model, cur_class)
if opt.reg_lambda != 0:
update_ewc_data(model, model.ewcData, model.synData, 0.001, 1)
if rank in [-1, 0]:
# Strip optimizers
n = opt.name if opt.name.isnumeric() else ''
fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt'
for f1, f2 in zip(
[wdir / 'last.pt', wdir / 'best.pt', results_file],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
if str(f2).endswith('.pt'): # is *.pt
strip_optimizer(f2) # strip optimizer
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket else None # upload
# Finish
plot_results(save_dir=log_dir) # save as results.png
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
#model=ema.ema,
model=model,
single_cls=opt.single_cls,
dataloader=all_test_dataloader,
save_dir=log_dir,
#plots=epoch == 0 or final_epoch, # plot first and last
log_imgs=opt.log_imgs,
verbose=True)
#wandb.log({'per class/AP per class All': maps[0]})
#tb_writer.add_scalar('per class/AP per class All', maps[0])
# Log
tags = [ # train loss
'test/precision', 'test/recall', 'test/mAP_0.5',
'test/mAP_0.5:0.95', 'test/giou_loss', 'test/obj_loss',
'test/cls_loss'
] # params
for x, tag in zip(list(results), tags):
if tb_writer:
tb_writer.add_scalar(tag, x, core_batch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
return results
def train(hyp):
epochs = opt.epochs # 300
batch_size = 18 # 64
weights = opt.weights # initial training weights
# Configure
init_seeds(1)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
test_path = data_dict['val']
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg).to(device)
assert model.md['nc'] == nc, '%s nc=%g classes but %s nc=%g classes' % (opt.data, nc, opt.cfg, model.md['nc'])
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
if any(x % gs != 0 for x in opt.img_size):
print('WARNING: --img-size %g,%g must be multiple of %s max stride %g' % (*opt.img_size, opt.cfg, gs))
imgsz, imgsz_test = [make_divisible(x, gs) for x in opt.img_size] # image sizes (train, test)
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
optimizer = optim.Adam(pg0, lr=hyp['lr0']) if opt.adam else \
optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Load Model
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
ckpt['model'] = \
{k: v for k, v in ckpt['model'].state_dict().items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(ckpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s." \
% (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
start_epoch = ckpt['epoch'] + 1
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.9 + 0.1 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # do not move
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
if device.type != 'cpu' and torch.cuda.device_count() > 1:
model = nn.DataParallel(model, device_ids=[i for i in range(torch.cuda.device_count())], dim=0)
import torch
from utils.torch_utils import select_device, intersect_dicts
from utils.general import make_divisible, check_file, set_logging
from models.yolo import Model
model_path = 'adpt/train/mafia/pre-train-1024/weights/best.pt'
bs = 1
device = select_device('', batch_size=bs)
ckpt = torch.load(model_path, map_location=device)
print(ckpt.keys())
print(ckpt['model'].yaml)
model = Model(ckpt['model'].yaml, ch=3, nc=3).to(device) # create
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=['anchor']) # intersect
model.load_state_dict(state_dict, strict=False) # load
def load_checkpoint(type_,
weights,
device,
cfg=None,
hyp=None,
nc=None,
recipe=None,
resume=None,
rank=-1):
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
start_epoch = ckpt['epoch'] + 1 if 'epoch' in ckpt else 0
pickled = isinstance(ckpt['model'], nn.Module)
if pickled and type_ == 'ensemble':
# load ensemble using pickled
cfg = None
model = attempt_load(weights, map_location=device) # load FP32 model
state_dict = model.state_dict()
else:
# load model from config and weights
cfg = cfg or (ckpt['yaml'] if 'yaml' in ckpt else None) or \
(ckpt['model'].yaml if pickled else None)
model = Model(cfg,
ch=3,
nc=ckpt['nc'] if ('nc' in ckpt and not nc) else nc,
anchors=hyp.get('anchors') if hyp else None).to(device)
model_key = 'ema' if (type_ in ['ema', 'ensemble'] and 'ema' in ckpt
and ckpt['ema']) else 'model'
state_dict = ckpt[model_key].float().state_dict(
) if pickled else ckpt[model_key]
# turn gradients for params back on in case they were removed
for p in model.parameters():
p.requires_grad = True
# load sparseml recipe for applying pruning and quantization
recipe = recipe or (ckpt['recipe'] if 'recipe' in ckpt else None)
sparseml_wrapper = SparseMLWrapper(model, recipe)
if type_ in ['ema', 'ensemble']:
# apply the recipe to create the final state of the model when not training
sparseml_wrapper.apply()
else:
# intialize the recipe for training
sparseml_wrapper.initialize(start_epoch)
if type_ == 'train':
# load any missing weights from the model
exclude = [
'anchor'
] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=type_ != 'train') # load
model.float()
report = 'Transferred %g/%g items from %s' % (
len(state_dict), len(model.state_dict()), weights)
return model, {
'ckpt': ckpt,
'state_dict': state_dict,
'start_epoch': start_epoch,
'sparseml_wrapper': sparseml_wrapper,
'report': report,
}
def train(hyp, tb_writer, opt, device):
print(f'Hyperparameters {hyp}')
log_dir = tb_writer.log_dir if tb_writer else 'runs/evolution' # run directory
wdir = str(Path(log_dir) / 'weights') + os.sep # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir + 'last.pt'
best = wdir + 'best.pt'
results_file = log_dir + os.sep + 'results.txt'
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.local_rank
# TODO: Init DDP logging. Only the first process is allowed to log.
# Since I see lots of print here, the logging configuration is skipped here. We may see repeated outputs.
# Save run settings
with open(Path(log_dir) / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(Path(log_dir) / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(
data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Remove previous results
if rank in [-1, 0]:
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg, nc=nc).to(device)
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
# default DDP implementation is slow for accumulation according to: https://pytorch.org/docs/stable/notes/ddp.html
# all-reduce operation is carried out during loss.backward().
# Thus, there would be redundant all-reduce communications in a accumulation procedure,
# which means, the result is still right but the training speed gets slower.
# TODO: If acceleration is needed, there is an implementation of allreduce_post_accumulation
# in https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/BERT/run_pretraining.py
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if hyp['optimizer'] == 'adam': # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Load Model
with torch_distributed_zero_first(rank):
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
exclude = ['anchor'] # exclude keys
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if k in model.state_dict() and not any(x in k for x in exclude)
and model.state_dict()[k].shape == v.shape
}
model.load_state_dict(ckpt['model'], strict=False)
print('Transferred %g/%g items from %s' %
(len(ckpt['model']), len(model.state_dict()), weights))
except KeyError as e:
s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \
"Please delete or update %s and try again, or use --weights '' to train from scratch." \
% (weights, opt.cfg, weights, weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# epochs
start_epoch = ckpt['epoch'] + 1
if epochs < start_epoch:
print(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O1',
verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: ((
(1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if rank in [-1, 0] and wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project=opt.wandb,
name=Path(log_dir).stem,
id=ckpt.get('wandb_id') if 'ckpt' in locals() else None)
loggers = {'wandb': wandb} # loggers dict
# DP mode
if device.type != 'cpu' and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and device.type != 'cpu' and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
print('Using SyncBatchNorm()')
# Exponential moving average
ema = torch_utils.ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if device.type != 'cpu' and rank != -1:
model = DDP(model, device_ids=[rank], output_device=rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
local_rank=rank,
world_size=opt.world_size)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Testloader
if rank in [-1, 0]:
# local_rank is set to -1. Because only the first process is expected to do evaluation.
testloader = create_dataloader(test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
augment=False,
cache=opt.cache_images,
rect=True,
local_rank=-1,
world_size=opt.world_size)[0]
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = names
# Class frequency
if rank in [-1, 0]:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1.
# model._initialize_biases(cf.to(device))
plot_labels(labels, save_dir=log_dir)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Check anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Start training
t0 = time.time()
nw = max(3 * nb,
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
scheduler.last_epoch = start_epoch - 1 # do not move
if rank in [0, -1]:
print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
print('Using %g dataloader workers' % dataloader.num_workers)
print('Starting training for %g epochs...' % epochs)
# torch.autograd.set_detect_anomaly(True)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
# When in DDP mode, the generated indices will be broadcasted to synchronize dataset.
if dataset.image_weights:
# Generate indices.
if rank in [-1, 0]:
w = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
image_weights = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=w)
dataset.indices = random.choices(
range(dataset.n), weights=image_weights,
k=dataset.n) # rand weighted idx
# Broadcast.
if rank != -1:
indices = torch.zeros([dataset.n], dtype=torch.int)
if rank == 0:
indices[:] = torch.from_tensor(dataset.indices,
dtype=torch.int)
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
if rank in [-1, 0]:
print(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj',
'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(
ni, xi,
[0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi,
[0.9, hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets.to(device),
model) # scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
if ema is not None:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_cached() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if ni < 3:
f = str(Path(log_dir) /
('train_batch%g.jpg' % ni)) # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# result = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
# if tb_writer and result is not None:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif ni == 10 and wandb:
wandb.log(
{
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in Path(log_dir).glob('train*.jpg')
if x.exists()
]
},
commit=False)
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
scheduler.step()
# Only the first process in DDP mode is allowed to log or save checkpoints.
if rank in [-1, 0]:
# mAP
if ema is not None:
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
save_json=final_epoch
and opt.data.endswith(os.sep + 'coco.yaml'),
model=ema.ema.module
if hasattr(ema.ema, 'module') else ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
plots=final_epoch,
log_imgs=16 if wandb else 0,
save_dir=log_dir)
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
tags = [
'train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss',
'val/cls_loss'
]
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch)
if wandb:
wandb.log({tag: x}, step=epoch,
commit=tag == tags[-1]) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1,
-1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema.module if hasattr(ema, 'module') else ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict(),
'wandb_id':
wandb_run.id if wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = ('_'
if len(opt.name) and not opt.name.isnumeric() else '') + opt.name
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (
f2, opt.bucket)) if opt.bucket and ispt else None # upload
# Finish
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
if wandb:
files = [
'results.png', 'confusion_matrix.png',
*[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]
]
wandb.log({
"Results": [
wandb.Image(str(Path(log_dir) / f), caption=f)
for f in files if (Path(log_dir) / f).exists()
]
})
try:
print("last:", last)
wandb.log_artifact(artifact_or_path=str(last),
type='model',
name="last")
except ValueError:
print("last model not found in", last)
try:
print("flast:", flast)
wandb.log_artifact(artifact_or_path=str(flast),
type='model',
name="flast")
except ValueError:
print("flast model not found in", flast)
try:
print("best:", best)
wandb.log_artifact(artifact_or_path=str(best),
type='model',
name="best")
except ValueError:
print("best model not found in", best)
try:
print("fbest:", fbest)
wandb.log_artifact(artifact_or_path=str(fbest),
type='model',
name="fbest")
except ValueError:
print("fbest model not found in", fbest)
print('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if rank not in [-1, 0] else None
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None):
logger.info(
colorstr('hyperparameters: ') + ', '.join(f'{k}={v}'
for k, v in hyp.items()))
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.safe_dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.safe_dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.safe_load(f) # data dict
# Logging- Doing this before checking the dataset. Might update data_dict
loggers = {'wandb': None} # loggers dict
if rank in [-1, 0]:
opt.hyp = hyp # add hyperparameters
run_id = torch.load(weights).get('wandb_id') if weights.endswith(
'.pt') and os.path.isfile(weights) else None
wandb_logger = WandbLogger(opt, save_dir.stem, run_id, data_dict)
loggers['wandb'] = wandb_logger.wandb
data_dict = wandb_logger.data_dict
if wandb_logger.wandb:
weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp # WandbLogger might update weights, epochs if resuming
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
names = ['item'] if opt.single_cls and len(
data_dict['names']) != 1 else data_dict['names'] # class names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
## hyps : command line hyperparameters (overwrites hyp.yaml)
hyps = None
try:
if opt.hyps is not None and len(opt.hyps) > 0:
## hyps should evaluate to a python dict()
hyps = ast.literal_eval(opt.hyps)
## add hyps to hyp (overwrite)...
for k, v in hyps.items():
hyp[k] = v
except:
pfunc(f'ERROR: problem parsing hyps (hyperparameter string): {hyps}')
# Print swagger job json string...
if opt.job_str:
pfunc(f'swagger job submitted:\n{opt.job_str}\n')
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
model = Model(opt.cfg or ckpt['model'].yaml,
ch=3,
nc=nc,
anchors=hyp.get('anchors')).to(device) # create
exclude = [
'anchor'
] if (opt.cfg or hyp.get('anchors')) and not opt.resume else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc,
anchors=hyp.get('anchors')).to(device) # create
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
# Freeze
freeze = [] # parameter names to freeze (full or partial)
if hyp.get('freeze'): ## freeze backbone layers?
N = int(hyp['freeze']) + 1
freeze = ['model.%s.' % x for x in range(N)]
logger.info('Freezing first {} layers of network'.format(N))
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
logger.info('freezing %s' % k)
v.requires_grad = False
# ## create separate testing model
# test_model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create
# for k, v in test_model.named_parameters():
# v.requires_grad = False # freeze all layers
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lr_epochs, init_epochs = epochs, 0
if hyp.get('init_epochs'):
init_epochs = hyp['init_epochs']
lr_epochs += init_epochs
if opt.linear_lr:
lf = lambda x: (1 - x / (lr_epochs - 1)) * (1.0 - hyp['lrf']) + hyp[
'lrf'] # linear
else:
lf = one_cycle(1, hyp['lrf'], lr_epochs) # cosine 1->hyp['lrf']
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# EMA
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
ema.updates = ckpt['updates']
# Results
if ckpt.get('training_results') is not None:
results_file.write_text(
ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Fix Image sizes
gs = max(int(model.stride.max()), 32) # grid size (max stride)
nl = model.model[
-1].nl # number of detection layers (used for scaling hyp['obj'])
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Fix Crop size
if hyp.get('crop') and hyp['crop'] > 0:
hyp['crop'] = check_img_size(hyp['crop'], gs)
imgsz_test = hyp['crop']
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
pfunc('DOING DATA PARALLEL MODE!!!!!!!!!!!')
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# if rank in [-1, 0]:
pfunc(
f'RANK={rank} opt.world_size={opt.world_size} dist.get_rank()={dist.get_rank()} dist.get_world_size()={dist.get_world_size()}'
)
# Trainloader
trainloader, dataset = create_dataloader(
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images, # cache='disk',
cache_efficient_sampling=True,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '))
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(trainloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Testloader
# test_batch_size = batch_size
test_batch_size = 1 ## so test_batch_size-per-GPU = 1 (needed for DDP validation)
testloader = create_dataloader(test_path,
imgsz_test,
test_batch_size,
gs,
opt,
hyp=hyp,
cache=opt.cache_images and not opt.notest,
cache_efficient_sampling=True,
drop_last=False,
shuffle=False,
rect=True,
training=False,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5,
prefix=colorstr('val: '))[0]
# Process 0
new_best_model = False
if rank in [-1, 0]:
# orig_testloader = create_dataloader(test_path, imgsz_test, test_batch_size, gs, opt,
# hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1,
# world_size=opt.world_size, workers=opt.workers,
# # lazy_caching=True,
# pad=0.5, prefix=colorstr('val: '))[0]
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
if plots:
plot_labels(labels, names, save_dir, loggers)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
model.half().float() # pre-reduce anchor precision
# DDP mode
if cuda and rank != -1:
model = DDP(
model,
device_ids=[opt.local_rank],
output_device=opt.local_rank,
# nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698
find_unused_parameters=any(
isinstance(layer, nn.MultiheadAttention)
for layer in model.modules()))
# Model parameters
hyp['box'] *= 3. / nl # scale to layers
hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers
hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers
hyp['label_smoothing'] = opt.label_smoothing
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
# Start training
t0 = t1 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
compute_loss = ComputeLoss(model) # init loss class
###### resuming training run..... #########################################
if init_epochs > 0:
nw = -1
logger.info(
'Stepping lr_scheduler forward {} epochs...'.format(init_epochs))
for i in range(init_epochs - 1):
scheduler.step()
## check initial model performance....
# if init_epochs>0 and rank in [-1, 0]:
# test.test(opt.data,
# batch_size=test_batch_size,
# imgsz=imgsz_test,
# model=ema.ema,
# single_cls=opt.single_cls,
# dataloader=orig_testloader,
# save_dir=save_dir,
# verbose=True,
# plots=False,
# log_imgs=opt.log_imgs if wandb else 0,
# compute_loss=compute_loss)
###########################################################################
pfunc(f'Image sizes {imgsz} train, {imgsz_test} test\n'
f'Using {trainloader.num_workers} trainloader workers\n'
f'Logging results to {save_dir}\n'
f'Starting training for {epochs} epochs...')
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 / nc # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
trainloader.sampler.set_epoch(epoch)
pbar = enumerate(trainloader)
if rank in [-1, 0]:
t1 = time.time()
num_img = 0
steps = list(range(100, 0, -2))
pfunc(
'=========================================================================================================='
)
pfunc(f'Epoch {epoch+1}/{epochs}')
optimizer.zero_grad()
# logging.StreamHandler.terminator = ""
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
## simple progress indicator....
if rank in [-1, 0]:
prog = int(np.ceil(100 * (i + 1) / nb))
while len(steps) > 0 and prog >= steps[-1]:
step = steps.pop()
# pfunc('.')
if step % 10 == 0:
pfunc(f' {step}%')
# gpu_stats()
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device)) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
td = time.time() - t1
num_img += imgs.shape[0]
imgs_sec = (num_img / td) * opt.world_size
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
# Plot
if plots and ni < 5:
f = save_dir / f'train_batch{ni}.jpg' # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# if tb_writer:
# tb_writer.add_graph(torch.jit.trace(de_parallel(model), imgs, strict=False), []) # model graph
# # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
elif plots and ni == 10 and wandb_logger.wandb:
wandb_logger.log({
"Mosaics": [
wandb_logger.wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg') if x.exists()
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# logging.StreamHandler.terminator = "\n"
if rank in [-1, 0]:
pfunc(
(' ' + '%10s' * 3) % ('total_min', 'gpu_mem', 'imgs_sec'))
pfunc((' ' + '%10.2f' + '%10s' + '%10.4g') %
(((time.time() - t1) / 60), mem, imgs_sec))
t1 = time.time()
final_epoch = epoch + 1 == epochs
##################################################################################
## DDP VALIDATION....
# results = (mp, mr, mf1, map50, map)#, *(loss.cpu() / len(dataloader)).tolist())
try:
results = test_ddp(
opt,
de_parallel(model),
testloader,
rank,
device,
names,
)
if rank in [-1, 0]:
pfunc(f'Validation Time: {(time.time()-t1)/60:0.2f} min')
# Logging
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/F1',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
# 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb_logger.wandb:
wandb_logger.log({tag: x}) # W&B
# Update best fitness
# fitness = weighted combination of [P, R, F1, [email protected], [email protected]:0.95]
fi = fitness(np.array(results).reshape(1, -1))
if fi > best_fitness:
best_fitness = fi
wandb_logger.end_epoch(best_result=best_fitness == fi)
# Save model
if (not opt.nosave) or (final_epoch
and not opt.evolve): # if save
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
# 'training_results': results_file.read_text(),
'model':
deepcopy(de_parallel(model)).half(),
'ema':
deepcopy(ema.ema).half(),
'updates':
ema.updates,
'optimizer':
optimizer.state_dict(),
'wandb_id':
wandb_logger.wandb_run.id
if wandb_logger.wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
pfunc('Saving best model!')
torch.save(ckpt, best)
new_best_model = True
best_model_msg = f'Best Model: Epoch {epoch+1}, mF1={results[2]:0.3f}, [email protected]:0.95={results[4]:0.3f}'
if wandb_logger.wandb:
if ((epoch + 1) % opt.save_period == 0
and not final_epoch) and opt.save_period != -1:
wandb_logger.log_model(
last.parent,
opt,
epoch,
fi,
best_model=best_fitness == fi)
del ckpt
# Upload best model to s3
if (epoch + 1) % 10 == 0 and epoch > 15:
if new_best_model:
strip_optimizer(best)
upload_model(opt)
new_best_model = False
# print best model so far
pfunc(best_model_msg)
# Upload output log to s3
upload_log(opt)
## END DDP VALIDATION
except Exception as e:
pfunc('Validation failed.')
################################################################
# end epoch ----------------------------------------------------------------------------------------------------
# end training =====================================================================================================
if rank in [-1, 0]:
# ## Plots
# if plots:
# plot_results(save_dir=save_dir) # save as results.png
# if wandb_logger.wandb:
# files = ['results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]]
# wandb_logger.log({"Results": [wandb_logger.wandb.Image(str(save_dir / f), caption=f) for f in files
# if (save_dir / f).exists()]})
# ## Test best.pt
# pfunc('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
# # if opt.data.endswith('coco.yaml') and nc == 80: # if COCO
# for m in [best] if best.exists() else [last]: # speed, mAP tests
# results, _, _ = test.test(data_dict,
# batch_size=test_batch_size,
# imgsz=imgsz_test,
# model=attempt_load(m, device),#.half(),
# single_cls=opt.single_cls,
# dataloader=orig_testloader,
# save_dir=save_dir,
# # verbose=nc < 50 and final_epoch,
# # plots=plots and final_epoch,
# wandb_logger=wandb_logger,
# plots=False,
# # compute_loss=compute_loss,
# )
# ## Strip optimizers
# final = best if best.exists() else last # final model
# for f in last, best:
# if f.exists():
# strip_optimizer(f) # strip optimizers
# if opt.bucket:
# os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload
# if wandb_logger.wandb and not opt.evolve: # Log the stripped model
# wandb_logger.wandb.log_artifact(str(final), type='model',
# name='run_' + wandb_logger.wandb_run.id + '_model',
# aliases=['latest', 'best', 'stripped'])
wandb_logger.finish_run()
else:
dist.destroy_process_group()
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None, wandb=None):
logger.info(f"Hyperparameters {hyp}")
save_dir, epochs, batch_size, total_batch_size, weights, rank = (
Path(opt.save_dir),
opt.epochs,
opt.batch_size,
opt.total_batch_size,
opt.weights,
opt.global_rank,
)
# Directories
wdir = save_dir / "weights"
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / "last.pt"
best = wdir / "best.pt"
results_file = save_dir / "results.txt"
# Save run settings
with open(save_dir / "hyp.yaml", "w") as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / "opt.yaml", "w") as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != "cpu"
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict["train"]
test_path = data_dict["val"]
nc, names = (
(1, ["item"]) if opt.single_cls else (int(data_dict["nc"]), data_dict["names"])
) # number classes, names
assert len(names) == nc, "%g names found for nc=%g dataset in %s" % (
len(names),
nc,
opt.data,
) # check
# Model
pretrained = weights.endswith(".pt")
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get("anchors"):
ckpt["model"].yaml["anchors"] = round(hyp["anchors"]) # force autoanchor
model = Model(opt.cfg or ckpt["model"].yaml, ch=3, nc=nc).to(device) # create
exclude = ["anchor"] if opt.cfg or hyp.get("anchors") else [] # exclude keys
state_dict = ckpt["model"].float().state_dict() # to FP32
state_dict = intersect_dicts(
state_dict, model.state_dict(), exclude=exclude
) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
"Transferred %g/%g items from %s"
% (len(state_dict), len(model.state_dict()), weights)
) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print("freezing %s" % k)
v.requires_grad = False
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(
round(nbs / total_batch_size), 1
) # accumulate loss before optimizing
hyp["weight_decay"] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, "bias") and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, "weight") and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(
pg0, lr=hyp["lr0"], betas=(hyp["momentum"], 0.999)
) # adjust beta1 to momentum
else:
optimizer = optim.SGD(
pg0, lr=hyp["lr0"], momentum=hyp["momentum"], nesterov=True
)
optimizer.add_param_group(
{"params": pg1, "weight_decay": hyp["weight_decay"]}
) # add pg1 with weight_decay
optimizer.add_param_group({"params": pg2}) # add pg2 (biases)
logger.info(
"Optimizer groups: %g .bias, %g conv.weight, %g other"
% (len(pg2), len(pg1), len(pg0))
)
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = (
lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp["lrf"])
+ hyp["lrf"]
) # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project="YOLOv5" if opt.project == "runs/train" else Path(opt.project).stem,
name=save_dir.stem,
id=ckpt.get("wandb_id") if "ckpt" in locals() else None,
)
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt["optimizer"] is not None:
optimizer.load_state_dict(ckpt["optimizer"])
best_fitness = ckpt["best_fitness"]
# Results
if ckpt.get("training_results") is not None:
with open(results_file, "w") as file:
file.write(ckpt["training_results"]) # write results.txt
# Epochs
start_epoch = ckpt["epoch"] + 1
if opt.resume:
assert (
start_epoch > 0
), "%s training to %g epochs is finished, nothing to resume." % (
weights,
epochs,
)
if epochs < start_epoch:
logger.info(
"%s has been trained for %g epochs. Fine-tuning for %g additional epochs."
% (weights, ckpt["epoch"], epochs)
)
epochs += ckpt["epoch"] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [
check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info("Using SyncBatchNorm()")
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if cuda and rank != -1:
model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert (
mlc < nc
), "Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g" % (
mlc,
nc,
opt.data,
nc - 1,
)
# Process 0
if rank in [-1, 0]:
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(
test_path,
imgsz_test,
total_batch_size,
gs,
opt,
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
)[
0
] # testloader
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
plot_labels(labels, save_dir=save_dir)
if tb_writer:
tb_writer.add_histogram("classes", c, 0)
if wandb:
wandb.log(
{
"Labels": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob("*labels*.png")
]
}
)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset, model=model, thr=hyp["anchor_t"], imgsz=imgsz)
# Model parameters
hyp["cls"] *= nc / 80.0 # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device
) # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(
round(hyp["warmup_epochs"] * nb), 1000
) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
logger.info(
"Image sizes %g train, %g test\n"
"Using %g dataloader workers\nLogging results to %s\n"
"Starting training for %g epochs..."
% (imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs)
)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = (
model.class_weights.cpu().numpy() * (1 - maps) ** 2
) # class weights
iw = labels_to_image_weights(
dataset.labels, nc=nc, class_weights=cw
) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw, k=dataset.n
) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (
torch.tensor(dataset.indices)
if rank == 0
else torch.zeros(dataset.n)
).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
("\n" + "%10s" * 8)
% ("Epoch", "gpu_mem", "box", "obj", "cls", "total", "targets", "img_size")
)
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs,
targets,
paths,
_,
) in (
pbar
): # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = (
imgs.to(device, non_blocking=True).float() / 255.0
) # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()
)
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x["lr"] = np.interp(
ni,
xi,
[
hyp["warmup_bias_lr"] if j == 2 else 0.0,
x["initial_lr"] * lf(epoch),
],
)
if "momentum" in x:
x["momentum"] = np.interp(
ni, xi, [hyp["warmup_momentum"], hyp["momentum"]]
)
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [
math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(
imgs, size=ns, mode="bilinear", align_corners=False
)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device), model
) # loss scaled by batch_size
if rank != -1:
loss *= (
opt.world_size
) # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = "%.3gG" % (
torch.cuda.memory_reserved() / 1e9
if torch.cuda.is_available()
else 0
) # (GB)
s = ("%10s" * 2 + "%10.4g" * 6) % (
"%g/%g" % (epoch, epochs - 1),
mem,
*mloss,
targets.shape[0],
imgs.shape[-1],
)
pbar.set_description(s)
# Plot
if plots and ni < 3:
f = save_dir / f"train_batch{ni}.jpg" # filename
plot_images(images=imgs, targets=targets, paths=paths, fname=f)
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif plots and ni == 3 and wandb:
wandb.log(
{
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob("train*.jpg")
]
}
)
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x["lr"] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
ema.update_attr(
model, include=["yaml", "nc", "hyp", "gr", "names", "stride"]
)
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch,
log_imgs=opt.log_imgs if wandb else 0,
)
# Write
with open(results_file, "a") as f:
f.write(
s + "%10.4g" * 7 % results + "\n"
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system(
"gsutil cp %s gs://%s/results/results%s.txt"
% (results_file, opt.bucket, opt.name)
)
# Log
tags = [
"train/box_loss",
"train/obj_loss",
"train/cls_loss", # train loss
"metrics/precision",
"metrics/recall",
"metrics/mAP_0.5",
"metrics/mAP_0.5:0.95",
"val/box_loss",
"val/obj_loss",
"val/cls_loss", # val loss
"x/lr0",
"x/lr1",
"x/lr2",
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
# Update best mAP
fi = fitness(
np.array(results).reshape(1, -1)
) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, "r") as f: # create checkpoint
ckpt = {
"epoch": epoch,
"best_fitness": best_fitness,
"training_results": f.read(),
"model": ema.ema,
"optimizer": None if final_epoch else optimizer.state_dict(),
"wandb_id": wandb_run.id if wandb else None,
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = opt.name if opt.name.isnumeric() else ""
fresults, flast, fbest = (
save_dir / f"results{n}.txt",
wdir / f"last{n}.pt",
wdir / f"best{n}.pt",
)
for f1, f2 in zip(
[wdir / "last.pt", wdir / "best.pt", results_file], [flast, fbest, fresults]
):
if f1.exists():
os.rename(f1, f2) # rename
if str(f2).endswith(".pt"): # is *.pt
strip_optimizer(f2) # strip optimizer
os.system(
"gsutil cp %s gs://%s/weights" % (f2, opt.bucket)
) if opt.bucket else None # upload
# Finish
if plots:
plot_results(save_dir=save_dir) # save as results.png
if wandb:
files = [
"results.png",
"precision_recall_curve.png",
"confusion_matrix.png",
]
wandb.log(
{
"Results": [
wandb.Image(str(save_dir / f), caption=f)
for f in files
if (save_dir / f).exists()
]
}
)
logger.info(
"%g epochs completed in %.3f hours.\n"
% (epoch - start_epoch + 1, (time.time() - t0) / 3600)
)
else:
dist.destroy_process_group()
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def _create(name,
pretrained=True,
channels=3,
classes=80,
autoshape=True,
verbose=True,
device=None):
"""Creates a specified YOLOv5 model
Arguments:
name (str): name of model, i.e. 'yolov5s'
pretrained (bool): load pretrained weights into the model
channels (int): number of input channels
classes (int): number of model classes
autoshape (bool): apply YOLOv5 .autoshape() wrapper to model
verbose (bool): print all information to screen
device (str, torch.device, None): device to use for model parameters
Returns:
YOLOv5 pytorch model
"""
from pathlib import Path
from models.yolo import Model, attempt_load
from utils.general import check_requirements, set_logging
from utils.google_utils import attempt_download
from utils.torch_utils import select_device
check_requirements(requirements=Path(__file__).parent / 'requirements.txt',
exclude=('tensorboard', 'pycocotools', 'thop',
'opencv-python'))
set_logging(verbose=verbose)
fname = Path(name).with_suffix('.pt') # checkpoint filename
try:
if pretrained and channels == 3 and classes == 80:
model = attempt_load(
fname,
map_location=torch.device('cpu')) # download/load FP32 model
else:
cfg = list((Path(__file__).parent /
'models').rglob(f'{name}.yaml'))[0] # model.yaml path
model = Model(cfg, channels, classes) # create model
if pretrained:
attempt_download(fname) # download if not found locally
ckpt = torch.load(fname,
map_location=torch.device('cpu')) # load
msd = model.state_dict() # model state_dict
csd = ckpt['model'].float().state_dict(
) # checkpoint state_dict as FP32
csd = {
k: v
for k, v in csd.items() if msd[k].shape == v.shape
} # filter
model.load_state_dict(csd, strict=False) # load
if len(ckpt['model'].names) == classes:
model.names = ckpt[
'model'].names # set class names attribute
if autoshape:
model = model.autoshape() # for file/URI/PIL/cv2/np inputs and NMS
device = select_device('0' if torch.cuda.is_available() else 'cpu'
) if device is None else torch.device(device)
return model.to(device)
except Exception as e:
help_url = 'https://github.com/ultralytics/yolov5/issues/36'
s = 'Cache may be out of date, try `force_reload=True`. See %s for help.' % help_url
raise Exception(s) from e
def test(data,
weights=None,
batch_size=16,
imgsz=640,
conf_thres=0.001,
iou_thres=0.6, # for NMS
save_json=False,
single_cls=False,
augment=False,
verbose=False,
model=None,
dataloader=None,
fast=False):
# Initialize/load model and set device
if model is None:
training = False
device = torch_utils.select_device(opt.device, batch_size=batch_size)
half = device.type != 'cpu' # half precision only supported on CUDA
# Remove previous
for f in glob.glob('test_batch*.jpg'):
os.remove(f)
# Load model
google_utils.attempt_download(weights)
# model = torch.load(weights, map_location=device)['state_dict'].float() # load to FP32
model = Model(model_cfg='/home/ai/yulu/yolov5/models/yolov5s.yaml')
if os.path.exists(opt.weights):
ckpt = torch.load(opt.weights, map_location=device)
state_dict = {key: ckpt['state_dict'][key] for key in model.state_dict().keys()}
model.load_state_dict(state_dict)
# torch_utils.model_info(model)
model.fuse()
model.to(device)
if half:
model.half() # to FP16
# Multi-GPU disabled, incompatible with .half()
# if device.type != 'cpu' and torch.cuda.device_count() > 1:
# model = nn.DataParallel(model)
else: # called by train.py
training = True
device = next(model.parameters()).device # get model device
# half disabled https://github.com/ultralytics/yolov5/issues/99
half = False # device.type != 'cpu' and torch.cuda.device_count() == 1
if half:
model.half() # to FP16
# Configure
model.eval()
with open(data) as f:
data = yaml.load(f, Loader=yaml.FullLoader) # model dict
nc = 1 if single_cls else int(data['nc']) # number of classes
iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95
# iouv = iouv[0].view(1) # comment for [email protected]:0.95
niou = iouv.numel()
# Dataloader
if dataloader is None: # not training
img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img
_ = model(img.half() if half else img) if device.type != 'cpu' else None # run once
fast |= conf_thres > 0.001 # enable fast mode
path = data['test'] if opt.task == 'test' else data['val'] # path to val/test images
dataset = LoadImagesAndLabels(path,
imgsz,
batch_size,
rect=True, # rectangular inference
single_cls=opt.single_cls, # single class mode
pad=0.5) # padding
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers
dataloader = DataLoader(dataset,
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
seen = 0
# names = model.names if hasattr(model, 'names') else model.module.names
coco91class = coco80_to_coco91_class()
s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95')
p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0.
loss = torch.zeros(3, device=device)
jdict, stats, ap, ap_class = [], [], [], []
for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)):
img = img.to(device)
img = img.half() if half else img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
nb, _, height, width = img.shape # batch size, channels, height, width
whwh = torch.Tensor([width, height, width, height]).to(device)
# Disable gradients
with torch.no_grad():
# Run model
t = torch_utils.time_synchronized()
inf_out, train_out = model(img, augment=augment) # inference and training outputs
t0 += torch_utils.time_synchronized() - t
# Compute loss
if training: # if model has loss hyperparameters
loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # GIoU, obj, cls
# Run NMS
t = torch_utils.time_synchronized()
output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, fast=fast) # ??????????????
t1 += torch_utils.time_synchronized() - t
# Statistics per image
for si, pred in enumerate(output):
labels = targets[targets[:, 0] == si, 1:]
nl = len(labels)
tcls = labels[:, 0].tolist() if nl else [] # target class
seen += 1
if pred is None:
if nl:
stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls))
continue
# Append to text file
# with open('test.txt', 'a') as file:
# [file.write('%11.5g' * 7 % tuple(x) + '\n') for x in pred]
# Clip boxes to image bounds
clip_coords(pred, (height, width))
# Append to pycocotools JSON dictionary
if save_json:
# [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ...
image_id = int(Path(paths[si]).stem.split('_')[-1])
box = pred[:, :4].clone() # xyxy
scale_coords(img[si].shape[1:], box, shapes[si][0], shapes[si][1]) # to original shape
box = xyxy2xywh(box) # xywh
box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner
for p, b in zip(pred.tolist(), box.tolist()):
jdict.append({'image_id': image_id,
'category_id': coco91class[int(p[5])],
'bbox': [round(x, 3) for x in b],
'score': round(p[4], 5)})
# Assign all predictions as incorrect
correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device)
if nl:
detected = [] # target indices
tcls_tensor = labels[:, 0]
# target boxes
tbox = xywh2xyxy(labels[:, 1:5]) * whwh
# Per target class
for cls in torch.unique(tcls_tensor):
ti = (cls == tcls_tensor).nonzero().view(-1) # prediction indices
pi = (cls == pred[:, 5]).nonzero().view(-1) # target indices
# Search for detections
if pi.shape[0]:
# Prediction to target ious
ious, i = box_iou(pred[pi, :4], tbox[ti]).max(1) # best ious, indices
# Append detections
for j in (ious > iouv[0]).nonzero():
d = ti[i[j]] # detected target
if d not in detected:
detected.append(d)
correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn
if len(detected) == nl: # all targets already located in image
break
# Append statistics (correct, conf, pcls, tcls)
stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls))
# Plot images
if batch_i < 1:
f = 'test_batch%g_gt.jpg' % batch_i # filename
# plot_images(img, targets, paths, f, names) # ground truth
plot_images(img, targets, paths, f, 'yolov5s')
f = 'test_batch%g_pred.jpg' % batch_i
# plot_images(img, output_to_target(output, width, height), paths, f, names) # predictions
plot_images(img, output_to_target(output, width, height), paths, f, 'yolov5s')
# Compute statistics
stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy
if len(stats):
p, r, ap, f1, ap_class = ap_per_class(*stats)
p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1) # [P, R, [email protected], [email protected]:0.95]
mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class
else:
nt = torch.zeros(1)
# Print results
pf = '%20s' + '%12.3g' * 6 # print format
print(pf % ('all', seen, nt.sum(), mp, mr, map50, map))
# Print results per class
if verbose and nc > 1 and len(stats):
for i, c in enumerate(ap_class):
print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))
# Print speeds
t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple
if not training:
print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t)
# Save JSON
if save_json and map50 and len(jdict):
imgIds = [int(Path(x).stem.split('_')[-1]) for x in dataloader.dataset.img_files]
f = 'detections_val2017_%s_results.json' % \
(weights.split(os.sep)[-1].replace('.pt', '') if weights else '') # filename
print('\nCOCO mAP with pycocotools... saving %s...' % f)
with open(f, 'w') as file:
json.dump(jdict, file)
try:
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
# https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
cocoGt = COCO(glob.glob('../coco/annotations/instances_val*.json')[0]) # initialize COCO ground truth api
cocoDt = cocoGt.loadRes(f) # initialize COCO pred api
cocoEval = COCOeval(cocoGt, cocoDt, 'bbox')
cocoEval.params.imgIds = imgIds # image IDs to evaluate
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
map, map50 = cocoEval.stats[:2] # update results ([email protected]:0.95, [email protected])
except:
print('WARNING: pycocotools must be installed with numpy==1.17 to run correctly. '
'See https://github.com/cocodataset/cocoapi/issues/356')
# Return results
maps = np.zeros(nc) + map
for i, c in enumerate(ap_class):
maps[c] = ap[i]
return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def detect():
source, weights, view_img, save_txt, imgsz = \
opt.input_images, opt.checkpoint, opt.view_img, opt.save_txt, opt.img_size
webcam = source == '0' or source.startswith('rtsp') or source.startswith('http') or source.endswith('.txt')
# Initialize
device = torch_utils.select_device(opt.device)
if os.path.exists(opt.output_images):
shutil.rmtree(opt.output_images) # delete output folder
os.makedirs(opt.output_images) # make new output folder
half = device.type != 'cpu' # half precision only supported on CUDA
with open(opt.project) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader)
model = Model(data_dict).to(device)
model.load_state_dict(torch.load(weights, map_location=device))
model.names = data_dict['names']
# Load model
#google_utils.attempt_download(weights)
#model = torch.load(weights, map_location=device)['model'].float() # load to FP32
#model = torch.load(weights, map_location=device).float() # load to FP32
# torch.save(torch.load(weights, map_location=device), weights) # update model if SourceChangeWarning
# model.fuse()
model.to(device).eval()
if half:
model.half() # to FP16
# Get names and colors
names = model.module.names if hasattr(model, 'module') else model.names
#colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))]
colors = get_all_colors(len(names))
# Run inference
t0 = time.time()
img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img
_ = model(img.half() if half else img) if device.type != 'cpu' else None # run once
#for path, img, im0s, vid_cap in dataset:
for f in os.listdir(source):
t1 = time.time()
path = os.path.join(source, f)
im0s = cv2.imread(path)
img = letterbox(im0s, new_shape=imgsz)[0]
img = img[:, :, ::-1].transpose(2, 0, 1)
img = np.ascontiguousarray(img)
img = torch.from_numpy(img).to(device)
img = img.half() if half else img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
if img.ndimension() == 3:
img = img.unsqueeze(0)
# Inference
#t1 = torch_utils.time_synchronized()
pred = model(img, augment=opt.augment)[0]
# Apply NMS
pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms)
t2 = torch_utils.time_synchronized()
# Process detections
for i, det in enumerate(pred): # detections per image
p, s, im0 = path, '', im0s
s += '%gx%g ' % img.shape[2:] # print string
gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh
if det is not None and len(det):
# Rescale boxes from img_size to im0 size
det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
# Print results
for c in det[:, -1].unique():
n = (det[:, -1] == c).sum() # detections per class
s += '%g %ss, ' % (n, names[int(c)]) # add to string
# Write results
for *xyxy, conf, cls in det:
label = '%s %.2f' % (names[int(cls)], conf)
xmin, ymin, xmax, ymax = xyxy
color = colors[int(cls)]
cv2.rectangle(im0, (xmin, ymin), (xmax, ymax), color=color, thickness=2)
cv2.putText(im0, label, (xmin, ymax), cv2.FONT_HERSHEY_SIMPLEX, 0.6, color, 1)
cv2.imwrite(os.path.join(opt.output_images, f), im0)
print('%s Done. (%.3fs)' % (path, time.time() - t1))
print('Done. (%.3fs)' % (time.time() - t0))
class JDETracker(object):
def __init__(self, opt, frame_rate=30):
self.opt = opt
self.model = Model(opt.cfg, nc=1)
ckpt = torch.load(opt.weights, map_location='cpu')
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if k in self.model.state_dict()
and self.model.state_dict()[k].shape == v.shape
}
self.model.load_state_dict(ckpt['model'], strict=False)
self.model.cuda().half().eval()
self.tracked_stracks = [] # type: list[STrack]
self.lost_stracks = [] # type: list[STrack]
self.removed_stracks = [] # type: list[STrack]
self.frame_id = 0
self.det_thresh = opt.conf_thres
self.buffer_size = int(frame_rate / 30.0 * opt.track_buffer)
self.max_time_lost = self.buffer_size
self.kalman_filter = KalmanFilter()
def update(self, im_blob, img0, path):
self.frame_id += 1
activated_starcks = [
] # for storing active tracks, for the current frame
refind_stracks = [
] # Lost Tracks whose detections are obtained in the current frame
lost_stracks = [
] # The tracks which are not obtained in the current frame but are not removed.(Lost for some time lesser than the threshold for removing)
removed_stracks = []
t1 = time.time()
''' Step 1: Network forward, get detections & embeddings'''
with torch.no_grad():
pred = self.model(im_blob)[0]
# pred is tensor of all the proposals (default number of proposals: 54264). Proposals have information associated with the bounding box and embeddings
pred = pred[pred[:, :, 4] > self.opt.conf_thres]
# pred now has lesser number of proposals. Proposals rejected on basis of object confidence score
if len(pred) > 0:
dets = non_max_suppression(pred.unsqueeze(0), self.opt.conf_thres,
self.opt.nms_thres)[0].cpu()
# Final proposals are obtained in dets. Information of bounding box and embeddings also included
# Next step changes the detection scales
scale_coords(self.opt.img_size, dets[:, :4], img0.shape).round()
'''Detections is list of (x1, y1, x2, y2, object_conf, class_score, class_pred)'''
# class_pred is the embeddings.
detections = [
STrack(STrack.tlbr_to_tlwh(tlbrs[:4]), tlbrs[4], f.numpy(), 30)
for (tlbrs, f) in zip(dets[:, :5], dets[:, 6:])
]
else:
detections = []
t2 = time.time()
# print('Forward: {} s'.format(t2-t1))
''' Add newly detected tracklets to tracked_stracks'''
unconfirmed = []
tracked_stracks = [] # type: list[STrack]
for track in self.tracked_stracks:
if not track.is_activated:
# previous tracks which are not active in the current frame are added in unconfirmed list
unconfirmed.append(track)
# print("Should not be here, in unconfirmed")
else:
# Active tracks are added to the local list 'tracked_stracks'
tracked_stracks.append(track)
''' Step 2: First association, with embedding'''
# Combining currently tracked_stracks and lost_stracks
strack_pool = joint_stracks(tracked_stracks, self.lost_stracks)
# Predict the current location with KF
STrack.multi_predict(strack_pool, self.kalman_filter)
dists = matching.embedding_distance(strack_pool, detections)
# dists = matching.gate_cost_matrix(self.kalman_filter, dists, strack_pool, detections)
dists = matching.fuse_motion(self.kalman_filter, dists, strack_pool,
detections)
# The dists is the list of distances of the detection with the tracks in strack_pool
matches, u_track, u_detection = matching.linear_assignment(dists,
thresh=0.7)
# The matches is the array for corresponding matches of the detection with the corresponding strack_pool
for itracked, idet in matches:
# itracked is the id of the track and idet is the detection
track = strack_pool[itracked]
det = detections[idet]
if track.state == TrackState.Tracked:
# If the track is active, add the detection to the track
track.update(detections[idet], self.frame_id)
activated_starcks.append(track)
else:
# We have obtained a detection from a track which is not active, hence put the track in refind_stracks list
track.re_activate(det, self.frame_id, new_id=False)
refind_stracks.append(track)
# None of the steps below happen if there are no undetected tracks.
''' Step 3: Second association, with IOU'''
detections = [detections[i] for i in u_detection]
# detections is now a list of the unmatched detections
r_tracked_stracks = [
] # This is container for stracks which were tracked till the
# previous frame but no detection was found for it in the current frame
for i in u_track:
if strack_pool[i].state == TrackState.Tracked:
r_tracked_stracks.append(strack_pool[i])
dists = matching.iou_distance(r_tracked_stracks, detections)
matches, u_track, u_detection = matching.linear_assignment(dists,
thresh=0.5)
# matches is the list of detections which matched with corresponding tracks by IOU distance method
for itracked, idet in matches:
track = r_tracked_stracks[itracked]
det = detections[idet]
if track.state == TrackState.Tracked:
track.update(det, self.frame_id)
activated_starcks.append(track)
else:
track.re_activate(det, self.frame_id, new_id=False)
refind_stracks.append(track)
# Same process done for some unmatched detections, but now considering IOU_distance as measure
for it in u_track:
track = r_tracked_stracks[it]
if not track.state == TrackState.Lost:
track.mark_lost()
lost_stracks.append(track)
# If no detections are obtained for tracks (u_track), the tracks are added to lost_tracks list and are marked lost
'''Deal with unconfirmed tracks, usually tracks with only one beginning frame'''
detections = [detections[i] for i in u_detection]
dists = matching.iou_distance(unconfirmed, detections)
matches, u_unconfirmed, u_detection = matching.linear_assignment(
dists, thresh=0.7)
for itracked, idet in matches:
unconfirmed[itracked].update(detections[idet], self.frame_id)
activated_starcks.append(unconfirmed[itracked])
# The tracks which are yet not matched
for it in u_unconfirmed:
track = unconfirmed[it]
track.mark_removed()
removed_stracks.append(track)
# after all these confirmation steps, if a new detection is found, it is initialized for a new track
""" Step 4: Init new stracks"""
for inew in u_detection:
track = detections[inew]
if track.score < self.det_thresh:
continue
track.activate(self.kalman_filter, self.frame_id)
activated_starcks.append(track)
""" Step 5: Update state"""
# If the tracks are lost for more frames than the threshold number, the tracks are removed.
for track in self.lost_stracks:
if self.frame_id - track.end_frame > self.max_time_lost:
track.mark_removed()
removed_stracks.append(track)
# print('Remained match {} s'.format(t4-t3))
# Update the self.tracked_stracks and self.lost_stracks using the updates in this step.
self.tracked_stracks = [
t for t in self.tracked_stracks if t.state == TrackState.Tracked
]
self.tracked_stracks = joint_stracks(self.tracked_stracks,
activated_starcks)
self.tracked_stracks = joint_stracks(self.tracked_stracks,
refind_stracks)
# self.lost_stracks = [t for t in self.lost_stracks if t.state == TrackState.Lost] # type: list[STrack]
self.lost_stracks = sub_stracks(self.lost_stracks,
self.tracked_stracks)
self.lost_stracks.extend(lost_stracks)
self.lost_stracks = sub_stracks(self.lost_stracks,
self.removed_stracks)
self.removed_stracks.extend(removed_stracks)
self.tracked_stracks, self.lost_stracks = remove_duplicate_stracks(
self.tracked_stracks, self.lost_stracks)
# get scores of lost tracks
output_stracks = [
track for track in self.tracked_stracks if track.is_activated
]
logger.debug('===========Frame {}=========='.format(self.frame_id))
logger.debug('Activated: {}'.format(
[track.track_id for track in activated_starcks]))
logger.debug('Refind: {}'.format(
[track.track_id for track in refind_stracks]))
logger.debug('Lost: {}'.format(
[track.track_id for track in lost_stracks]))
logger.debug('Removed: {}'.format(
[track.track_id for track in removed_stracks]))
# print('Final {} s'.format(t5-t4))
return output_stracks
def train(hyp, opt, device, tb_writer=None, wandb=None):
logger.info(f'Hyperparameters {hyp}')
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
names = ['item'] if opt.single_cls and len(
data_dict['names']) != 1 else data_dict['names'] # class names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
# with torch_distributed_zero_first(rank):
# attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if rank in [-1, 0] and wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project='YOLOv5'
if opt.project == 'runs/train' else Path(opt.project).stem,
name=save_dir.stem,
id=ckpt.get('wandb_id') if 'ckpt' in locals() else None)
loggers = {'wandb': wandb} # loggers dict
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = int(model.stride.max()) # grid size (max stride)
nl = model.model[
-1].nl # number of detection layers (used for scaling hyp['obj'])
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# TODO 将cfg添加到配置变量中
cfg_model = Darknet('cfg/yolov5s_v4.cfg',
(opt.img_size[0], opt.img_size[0])).to(device)
# cfg_model = Darknet('cfg/yolov5s_v3.cfg', (416, 416)).to(device)
copy_weight_v4(model, cfg_model)
# 剪枝操作 sr开启稀疏训练 prune 不同的剪枝策略
# 剪枝操作
if opt.prune == 1:
CBL_idx, _, prune_idx, shortcut_idx, _ = parse_module_defs2(
cfg_model.module_defs)
if opt.sr:
print('shortcut sparse training')
elif opt.prune == 0:
CBL_idx, _, prune_idx = parse_module_defs(cfg_model.module_defs)
if opt.sr:
print('normal sparse training ')
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(
test_path,
imgsz_test,
total_batch_size,
gs,
opt, # testloader
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5)[0]
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
plot_labels(labels, save_dir, loggers)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
hyp['cls'] *= nc / 80. # scale hyp['cls'] to class count
hyp['obj'] *= imgsz**2 / 640.**2 * 3. / nl # scale hyp['obj'] to image size and output layers
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
for idx in prune_idx:
bn_weights = gather_bn_weights(cfg_model.module_list, [idx])
tb_writer.add_histogram('before_train_perlayer_bn_weights/hist',
bn_weights.numpy(),
idx,
bins='doane')
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
logger.info('Image sizes %g train, %g test\n'
'Using %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs))
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 / nc # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
sr_flag = get_sr_flag(epoch, opt.sr)
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
# with amp.autocast(enabled=cuda):
# pred = model(imgs) # forward
# loss, loss_items = compute_loss(pred, targets.to(device), model) # loss scaled by batch_size
# if rank != -1:
# loss *= opt.world_size # gradient averaged between devices in DDP mode
# if opt.quad:
# loss *= 4.
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets.to(device),
model) # scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
# scaler.scale(loss).backward()
loss.backward()
idx2mask = None
# if opt.sr and opt.prune==1 and epoch > opt.epochs * 0.5:
# idx2mask = get_mask2(model, prune_idx, 0.85)
# copy_weight(model,cfg_model)
BNOptimizer.updateBN(sr_flag, cfg_model.module_list, opt.s,
prune_idx, epoch, idx2mask, opt)
# Optimize
if ni % accumulate == 0:
# scaler.step(optimizer) # optimizer.step
# scaler.update()
optimizer.step()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if plots and ni < 3:
f = save_dir / f'train_batch{ni}.jpg' # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif plots and ni == 3 and wandb:
wandb.log({
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg')
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
ema.update_attr(model,
include=[
'yaml', 'nc', 'hyp', 'gr', 'names',
'stride', 'class_weights'
])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch,
log_imgs=opt.log_imgs if wandb else 0)
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
#剪枝后bn层权重
bn_weights = gather_bn_weights(cfg_model.module_list, prune_idx)
tb_writer.add_histogram('bn_weights/hist',
bn_weights.numpy(),
epoch,
bins='doane')
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict(),
'wandb_id':
wandb_run.id if wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
for idx in prune_idx:
bn_weights = gather_bn_weights(cfg_model.module_list, [idx])
tb_writer.add_histogram('after_train_perlayer_bn_weights/hist',
bn_weights.numpy(),
idx,
bins='doane')
if rank in [-1, 0]:
# Strip optimizers
final = best if best.exists() else last # final model
for f in [last, best]:
if f.exists():
strip_optimizer(f) # strip optimizers
if opt.bucket:
os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload
# Plots
if plots:
plot_results(save_dir=save_dir) # save as results.png
if wandb:
files = [
'results.png', 'precision_recall_curve.png',
'confusion_matrix.png'
]
wandb.log({
"Results": [
wandb.Image(str(save_dir / f), caption=f)
for f in files if (save_dir / f).exists()
]
})
if opt.log_artifacts:
wandb.log_artifact(artifact_or_path=str(final),
type='model',
name=save_dir.stem)
# Test best.pt
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
if opt.data.endswith('coco.yaml') and nc == 80: # if COCO
for conf, iou, save_json in ([0.25, 0.45,
False], [0.001, 0.65,
True]): # speed, mAP tests
results, _, _ = test.test(opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
conf_thres=conf,
iou_thres=iou,
model=attempt_load(final,
device).half(),
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
save_json=save_json,
plots=False)
else:
dist.destroy_process_group()
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None, wandb=None):
logger.info(
colorstr('hyperparameters: ') + ', '.join(f'{k}={v}'
for k, v in hyp.items()))
print(f'Hyperparameters {hyp}')
"""
训练日志包括：权重、tensorboard文件、超参数hyp、设置的训练参数opt(也就是epochs,batch_size等),result.txt
result.txt包括: 占GPU内存、训练集的GIOU loss, objectness loss, classification loss, 总loss,
targets的数量, 输入图片分辨率, 准确率TP/(TP+FP),召回率TP/P ;
测试集的mAP50, [email protected]:0.95, GIOU loss, objectness loss, classification loss.
还会保存batch<3的ground truth
"""
# 获取保存路径、总轮次、批次、总批次(涉及到分布式训练)、权重、进程序号(主要用于分布式训练)
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
# 保存hyp和opt
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
# torch_distributed_zero_first同步所有进程
# check_dataset检查数据集，如果没找到数据集则下载数据集(仅适用于项目中自带的yaml文件数据集)
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
names = ['item'] if opt.single_cls and len(
data_dict['names']) != 1 else data_dict['names'] # class names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
# 加载模型，从google云盘中自动下载模型
# 但通常会下载失败，建议提前下载下来放进weights目录
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
# 加载检查点
ckpt = torch.load(weights, map_location=device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
"""
这里模型创建，可通过opt.cfg，也可通过ckpt['model'].yaml
这里的区别在于是否是resume，resume时会将opt.cfg设为空，则按照ckpt['model'].yaml创建模型
这也影响着下面是否除去anchor的key(也就是不加载anchor)，如果resume则不加载anchor
主要是因为保存的模型会保存anchors，有时候用户自定义了anchor之后，再resume，则原来基于coco数据集的anchor就会覆盖自己设定的anchor，
参考https://github.com/ultralytics/yolov5/issues/459
所以下面设置了intersect_dicts，该函数就是忽略掉exclude
"""
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
# 显示加载预训练权重的的键值对和创建模型的键值对
# 如果设置了resume，则会少加载两个键值对(anchors,anchor_grid)
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
"""
冻结模型层,设置冻结层名字即可
具体可以查看https://github.com/ultralytics/yolov5/issues/679
但作者不鼓励冻结层,因为他的实验当中显示冻结层不能获得更好的性能,参照:https://github.com/ultralytics/yolov5/pull/707
并且作者为了使得优化参数分组可以正常进行,在下面将所有参数的requires_grad设为了True
其实这里只是给一个freeze的示例
"""
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
"""
nbs为模拟的batch_size;
就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64，
也就是模型梯度累积了64/16=4(accumulate)次之后
再更新一次模型，变相的扩大了batch_size
"""
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
# 根据accumulate设置权重衰减系数
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")
# 将模型分成三组(weight、bn, bias, 其他所有参数)优化
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
# 选用优化器，并设置pg0组的优化方式
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
# 设置weight、bn的优化方式
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
# 设置biases的优化方式
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
# 打印优化信息
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# 设置cosine调度器，定义学习率衰减学习率衰减，这里为余弦退火方式进行衰减
# 就是根据以下公式lf,epoch和超参数hyp['lrf']进行衰减
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
# lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Logging
if rank in [-1, 0] and wandb and wandb.run is None:
opt.hyp = hyp # add hyperparameters
wandb_run = wandb.init(
config=opt,
resume="allow",
project='YOLOv5'
if opt.project == 'runs/train' else Path(opt.project).stem,
name=save_dir.stem,
id=ckpt.get('wandb_id') if 'ckpt' in locals() else None)
loggers = {'wandb': wandb} # loggers dict
# EMA
# 在深度学习中，经常会使用EMA（指数移动平均）这个方法对模型的参数做滑动平均，以求提高测试指标并增加模型鲁棒，如果GPU进程数大于1,则不创建
# Exponential moving average
ema = ModelEMA(model) if rank in [-1, 0] else None
# Resume
# best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, [email protected], [email protected]:0.95]再求和所得
# 根据best_fitness来保存best.pt
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
# 加载优化器与best_fitness
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# EMA
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
ema.updates = ckpt['updates']
# Results
# 加载训练结果result.txt
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs 加载训练的轮次
start_epoch = ckpt['epoch'] + 1
"""
如果resume,则备份权重
尽管目前resume能够近似100%成功的起作用了,参照:https://github.com/ultralytics/yolov5/pull/756
但为了防止resume时出现其他问题,把之前的权重覆盖了,所以这里进行备份,参照:https://github.com/ultralytics/yolov5/pull/765
"""
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
"""
如果新设置epochs小于加载的epoch，
则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数
"""
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# 获取模型最大步长和模型输入图片分辨率
gs = int(model.stride.max()) # grid size (max stride)
nl = model.model[
-1].nl # number of detection layers (used for scaling hyp['obj'])
# 检查训练和测试图片分辨率确保能够整除总步长gs
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
# 分布式训练,参照:https://github.com/ultralytics/yolov5/issues/475
# DataParallel模式,仅支持单机多卡
# rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式
# rank=-1且gpu数量=1时,不会进行分布式
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
# 使用跨卡同步BN
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# DDP mode
# 如果rank不等于-1,则使用DistributedDataParallel模式
# local_rank为gpu编号,rank为进程,例如rank=3，local_rank=0 表示第 3 个进程内的第 1 块 GPU。
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '))
"""
获取标签中最大的类别值，并于类别数作比较
如果小于类别数则表示有问题
"""
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
# 更新ema模型的updates参数,保持ema的平滑性
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(
test_path,
imgsz_test,
total_batch_size,
gs,
opt, # testloader
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5,
prefix=colorstr('val: '))[0]
if not opt.resume:
# 将所有样本的标签拼接到一起shape为(total, 1)，统计后做可视化
labels = np.concatenate(dataset.labels, 0)
# 获得所有样本的类别
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
# 根据上面的统计对所有样本的类别，中心点xy位置，长宽wh做可视化
plot_labels(labels, save_dir, loggers)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Check anchors
"""
计算默认锚点anchor与数据集标签框的长宽比值
标签的长h宽w与anchor的长h_a宽w_a的比值, 即h/h_a, w/w_a都要在(1/hyp['anchor_t'], hyp['anchor_t'])是可以接受的
如果标签框满足上面条件的数量小于总数的99%，则根据k-mean算法聚类新的锚点anchor
"""
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters # 根据自己数据集的类别数设置分类损失的系数
hyp['box'] *= 3. / nl # scale to layers
hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers
hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
# 根据labels初始化图片采样权重
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
"""
设置giou的值在objectness loss中做标签的系数, 使用代码如下
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)
这里model.gr=1，也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签
"""
# Start training
t0 = time.time()
# 获取热身训练的迭代次数
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
"""
设置学习率衰减所进行到的轮次，
目的是打断训练后，--resume接着训练也能正常的衔接之前的训练进行学习率衰减
"""
scheduler.last_epoch = start_epoch - 1 # do not move
# 通过torch1.6自带的api设置混合精度训练
scaler = amp.GradScaler(enabled=cuda)
"""
打印训练和测试输入图片分辨率
加载图片时调用的cpu进程数
从哪个epoch开始训练
"""
logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n'
f'Using {dataloader.num_workers} dataloader workers\n'
f'Logging results to {save_dir}\n'
f'Starting training for {epochs} epochs...')
# 加载图片权重（可选），定义进度条，设置偏差Burn-in，使用多尺度，前向传播，损失函数，反向传播，优化器，打印进度条，保存训练参数至tensorboard，计算mAP，保存结果到results.txt，保存模型（最好和最后）
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
"""
如果设置进行图片采样策略，
则根据前面初始化的图片采样权重model.class_weights以及maps配合每张图片包含的类别数
通过random.choices生成图片索引indices从而进行采样
"""
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 / nc # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
# 类平衡采样
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
# 如果是DDP模式,则广播采样策略
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
# 初始化训练时打印的平均损失信息
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
# DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子，
# 每次epoch不同，随机种子就不同
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'targets', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar,
total=nb) # progress bar tqdm 创建进度条，方便训练时 信息的展示
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
"""
热身训练(前nw次迭代)
在前nw次迭代中，根据以下方式选取accumulate和学习率
"""
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
"""
bias的学习率从0.1下降到基准学习率lr*lf(epoch)，其他的参数学习率从0增加到lr*lf(epoch)
lf为上面设置的余弦退火的衰减函数
"""
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
# 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# 混合精度
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
# 计算损失，包括分类损失，objectness损失，框的回归损失
# loss为总损失值，loss_items为一个元组，包含分类损失，objectness损失，框的回归损失和总损失
if (IS_Debug()):
#loss, loss_items = compute_loss(pred, targets.to(device), model, imgs) # loss scaled by batch_size
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
else:
loss, loss_items = compute_loss(
pred, targets.to(device),
model) # loss scaled by batch_size
if rank != -1:
# 平均不同gpu之间的梯度
loss *= opt.world_size # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize
# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema is not None:
ema.update(model)
# Print
if rank in [-1, 0]:
# 打印显存，进行的轮次，损失，target的数量和图片的size等信息
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
# 将前三次迭代batch的标签框在图片上画出来并保存
if plots and ni < 3:
f = save_dir / f'train_batch{ni}.jpg' # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
elif plots and ni == 3 and wandb:
wandb.log({
"Mosaics": [
wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg')
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
# 进行学习率衰减
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
# 更新EMA的属性
# 添加include的属性
if ema:
ema.update_attr(model,
include=[
'yaml', 'nc', 'hyp', 'gr', 'names',
'stride', 'class_weights'
])
# 判断该epoch是否为最后一轮
final_epoch = epoch + 1 == epochs
# 对测试集进行测试，计算mAP等指标
# 测试时使用的是EMA模型
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch,
log_imgs=opt.log_imgs if wandb else 0)
# Write
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 7 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb:
wandb.log({tag: x}) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
"""
保存模型，还保存了epoch，results，optimizer等信息，
optimizer将不会在最后一轮完成后保存
model保存的是EMA的模型
"""
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict(),
'wandb_id':
wandb_run.id if wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
"""
模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；
并且对模型进行model.half(), 将Float32的模型->Float16，
可以减少模型大小，提高inference速度
"""
final = best if best.exists() else last # final model
for f in [last, best]:
if f.exists():
strip_optimizer(f) # strip optimizers
if opt.bucket:
os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload
# Plots
if plots:
# 可视化results.txt文件
plot_results(save_dir=save_dir) # save as results.png
if wandb:
files = [
'results.png', 'precision_recall_curve.png',
'confusion_matrix.png'
]
wandb.log({
"Results": [
wandb.Image(str(save_dir / f), caption=f)
for f in files if (save_dir / f).exists()
]
})
if opt.log_artifacts:
wandb.log_artifact(artifact_or_path=str(final),
type='model',
name=save_dir.stem)
# Test best.pt
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
if opt.data.endswith('coco.yaml') and nc == 80: # if COCO
for conf, iou, save_json in ([0.25, 0.45,
False], [0.001, 0.65,
True]): # speed, mAP tests
results, _, _ = test.test(opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
conf_thres=conf,
iou_thres=iou,
model=attempt_load(final,
device).half(),
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
save_json=save_json,
plots=False)
else:
dist.destroy_process_group() # 释放显存
wandb.run.finish() if wandb and wandb.run else None
torch.cuda.empty_cache()
return results
def train(hyp):
epochs = opt.epochs # 300
batch_size = opt.batch_size # 64
weights = opt.weights # initial training weights
# Configure
init_seeds(1)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
test_path = data_dict['val']
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg).to(device)
assert model.md['nc'] == nc, '%s nc=%g classes but %s nc=%g classes' % (opt.data, nc, opt.cfg, model.md['nc'])
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
optimizer = optim.Adam(pg0, lr=hyp['lr0']) if opt.adam else \
optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Load Model
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items()
if model.state_dict()[k].shape == v.shape} # to FP32, filter
model.load_state_dict(ckpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s." \
% (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
start_epoch = ckpt['epoch'] + 1
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.9 + 0.1 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # do not move
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
dist.init_process_group(backend='nccl', # distributed backend
init_method='tcp://127.0.0.1:9999', # init method
world_size=1, # number of nodes
rank=0) # node rank
model = torch.nn.parallel.DistributedDataParallel(model)
# Dataset
dataset = LoadImagesAndLabels(train_path, imgsz, batch_size,
augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect, # rectangular training
cache_images=opt.cache_images,
single_cls=opt.single_cls)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (mlc, nc, opt.cfg)
# Dataloader
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=nw,
shuffle=not opt.rect, # Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
# Testloader
testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls),
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
def train(hyp, tb_writer, opt, device):
print(f'Hyperparameters {hyp}')
log_dir = tb_writer.log_dir if tb_writer else 'runs/evolution' # run directory
wdir = str(Path(log_dir) / 'weights') + os.sep # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir + 'last.pt'
best = wdir + 'best.pt'
results_file = log_dir + os.sep + 'results.txt'
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.local_rank
# TODO: Init DDP logging. Only the first process is allowed to log.
# Since I see lots of print here, the logging configuration is skipped here. We may see repeated outputs.
# Save run settings
with open(Path(log_dir) / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(Path(log_dir) / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
train_mask_path = data_dict['train_mask']
test_path = data_dict['val']
nc, names = (1, ['item']) if opt.single_cls else (int(data_dict['nc']), data_dict['names']) # number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data) # check
# Remove previous results
if rank in [-1, 0]:
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg, nc=nc).to(device)
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
# default DDP implementation is slow for accumulation according to: https://pytorch.org/docs/stable/notes/ddp.html
# all-reduce operation is carried out during loss.backward().
# Thus, there would be redundant all-reduce communications in a accumulation procedure,
# which means, the result is still right but the training speed gets slower.
# TODO: If acceleration is needed, there is an implementation of allreduce_post_accumulation
# in https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/BERT/run_pretraining.py
accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if hyp['optimizer'] == 'adam': # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Load Model
with torch_distributed_zero_first(rank):
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
exclude = ['anchor'] # exclude keys
ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items()
if k in model.state_dict() and not any(x in k for x in exclude)
and model.state_dict()[k].shape == v.shape}
model.load_state_dict(ckpt['model'], strict=False)
print('Transferred %g/%g items from %s' % (len(ckpt['model']), len(model.state_dict()), weights))
except KeyError as e:
s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \
"Please delete or update %s and try again, or use --weights '' to train from scratch." \
% (weights, opt.cfg, weights, weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# epochs
start_epoch = ckpt['epoch'] + 1
if epochs < start_epoch:
print('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %
(weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.8 + 0.2 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
# DP mode
if device.type != 'cpu' and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and device.type != 'cpu' and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
print('Using SyncBatchNorm()')
# Exponential moving average
ema = torch_utils.ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
if device.type != 'cpu' and rank != -1:
model = DDP(model, device_ids=[rank], output_device=rank)
# Trainloader
dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True,
cache=opt.cache_images, rect=opt.rect, local_rank=rank,
world_size=opt.world_size)
mask_dataloader, mask_dataset = create_dataloader(train_mask_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True,
cache=opt.cache_images, rect=opt.rect, local_rank=rank,
world_size=opt.world_size, multiple_augmentation=True)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1)
# Testloader
if rank in [-1, 0]:
# local_rank is set to -1. Because only the first process is expected to do evaluation.
testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False,
cache=opt.cache_images, rect=True, local_rank=-1, world_size=opt.world_size)[0]
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
model.names = names
# Class frequency
if rank in [-1, 0]:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1.
# model._initialize_biases(cf.to(device))
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Check anchors
if not opt.noautoanchor:
check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
# Start training
t0 = time.time()
nw = max(3 * nb, 1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
scheduler.last_epoch = start_epoch - 1 # do not move
if rank in [0, -1]:
print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
print('Using %g dataloader workers' % dataloader.num_workers)
print('Starting training for %g epochs...' % epochs)
# torch.autograd.set_detect_anomaly(True)
feature_loss = torch.nn.MSELoss()
teacher_model = attempt_load('./teacher.pt', fuse=False) # True
teacher_model = teacher_model.cuda()
dataset.epoch = start_epoch
mask_dataset.epoch = start_epoch
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
# When in DDP mode, the generated indices will be broadcasted to synchronize dataset.
if dataset.image_weights:
# Generate indices.
if rank in [-1, 0]:
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights,
k=dataset.n) # rand weighted idx
# Broadcast.
if rank != -1:
indices = torch.zeros([dataset.n], dtype=torch.int)
if rank == 0:
indices[:] = torch.from_tensor(dataset.indices, dtype=torch.int)
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
optimizer.zero_grad()
if rank != -1:
dataloader.sampler.set_epoch(epoch)
mask_dataloader.sampler.set_epoch(epoch)
train_data(model, teacher_model, dataloader, optimizer, ema,
device, rank, nw, nbs, total_batch_size
epoch=epoch, accumulate=accumulate, epochs=epochs)
s, mloss = train_data(model, teacher_model, mask_dataloader, optimizer, ema,
device, rank, nw, nbs, total_batch_size
epoch=epoch, accumulate=accumulate, epochs=epochs)
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
dataset.epoch += 1
mask_dataset.epoch += 1
scheduler.step()
# Only the first process in DDP mode is allowed to log or save checkpoints.
if rank in [-1, 0]:
# mAP
if ema is not None:
ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test_val.test(opt.data,
batch_size=total_batch_size,
imgsz=imgsz_test,
save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'),
model=ema.ema.module if hasattr(ema.ema, 'module') else ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=log_dir)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95',
'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': ema.ema.module if hasattr(ema, 'module') else ema.ema,
'optimizer': None if final_epoch else optimizer.state_dict()}
# Save last, best and delete
torch.save(ckpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.name
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket and ispt else None # upload
# Finish
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
return results
def train(
hyp, # path/to/hyp.yaml or hyp dictionary
opt,
device,
callbacks):
save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
# Directories
w = save_dir / 'weights' # weights dir
(w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir
last, best = w / 'last.pt', w / 'best.pt'
# Hyperparameters
if isinstance(hyp, str):
with open(hyp, errors='ignore') as f:
hyp = yaml.safe_load(f) # load hyps dict
LOGGER.info(
colorstr('hyperparameters: ') + ', '.join(f'{k}={v}'
for k, v in hyp.items()))
# Save run settings
if not evolve:
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.safe_dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.safe_dump(vars(opt), f, sort_keys=False)
# Loggers
data_dict = None
if RANK in [-1, 0]:
loggers = Loggers(save_dir, weights, opt, hyp,
LOGGER) # loggers instance
if loggers.wandb:
data_dict = loggers.wandb.data_dict
if resume:
weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
# Register actions
for k in methods(loggers):
callbacks.register_action(k, callback=getattr(loggers, k))
# Config
plots = not evolve # create plots
cuda = device.type != 'cpu'
init_seeds(1 + RANK)
with torch_distributed_zero_first(LOCAL_RANK):
data_dict = data_dict or check_dataset(data) # check if None
train_path, val_path = data_dict['train'], data_dict['val']
nc = 1 if single_cls else int(data_dict['nc']) # number of classes
names = ['item'] if single_cls and len(
data_dict['names']) != 1 else data_dict['names'] # class names
assert len(
names
) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check
is_coco = isinstance(val_path, str) and val_path.endswith(
'coco/val2017.txt') # COCO dataset
# Model
check_suffix(weights, '.pt') # check weights
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(LOCAL_RANK):
weights = attempt_download(
weights) # download if not found locally
ckpt = torch.load(weights, map_location='cpu'
) # load checkpoint to CPU to avoid CUDA memory leak
model = Model(cfg or ckpt['model'].yaml,
ch=3,
nc=nc,
anchors=hyp.get('anchors')).to(device) # create
exclude = [
'anchor'
] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys
csd = ckpt['model'].float().state_dict(
) # checkpoint state_dict as FP32
csd = intersect_dicts(csd, model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(csd, strict=False) # load
LOGGER.info(
f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}'
) # report
else:
model = Model(cfg, ch=3, nc=nc,
anchors=hyp.get('anchors')).to(device) # create
# Freeze
freeze = [
f'model.{x}.'
for x in (freeze if len(freeze) > 1 else range(freeze[0]))
] # layers to freeze
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
LOGGER.info(f'freezing {k}')
v.requires_grad = False
# Image size
gs = max(int(model.stride.max()), 32) # grid size (max stride)
imgsz = check_img_size(opt.imgsz, gs,
floor=gs * 2) # verify imgsz is gs-multiple
# Batch size
if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size
batch_size = check_train_batch_size(model, imgsz)
loggers.on_params_update({"batch_size": batch_size})
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}")
g0, g1, g2 = [], [], [] # optimizer parameter groups
for v in model.modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias
g2.append(v.bias)
if isinstance(v, nn.BatchNorm2d): # weight (no decay)
g0.append(v.weight)
elif hasattr(v, 'weight') and isinstance(
v.weight, nn.Parameter): # weight (with decay)
g1.append(v.weight)
if opt.optimizer == 'Adam':
optimizer = Adam(g0, lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
elif opt.optimizer == 'AdamW':
optimizer = AdamW(g0, lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = SGD(g0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': g1,
'weight_decay': hyp['weight_decay']
}) # add g1 with weight_decay
optimizer.add_param_group({'params': g2}) # add g2 (biases)
LOGGER.info(
f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups "
f"{len(g0)} weight (no decay), {len(g1)} weight, {len(g2)} bias")
del g0, g1, g2
# Scheduler
if opt.cos_lr:
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
else:
lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf'
] # linear
scheduler = lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs)
# EMA
ema = ModelEMA(model) if RANK in [-1, 0] else None
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# EMA
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
ema.updates = ckpt['updates']
# Epochs
start_epoch = ckpt['epoch'] + 1
if resume:
assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.'
if epochs < start_epoch:
LOGGER.info(
f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs."
)
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, csd
# DP mode
if cuda and RANK == -1 and torch.cuda.device_count() > 1:
LOGGER.warning(
'WARNING: DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n'
'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.'
)
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and RANK != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
LOGGER.info('Using SyncBatchNorm()')
# Trainloader
train_loader, dataset = create_dataloader(
train_path,
imgsz,
batch_size // WORLD_SIZE,
gs,
single_cls,
hyp=hyp,
augment=True,
cache=None if opt.cache == 'val' else opt.cache,
rect=opt.rect,
rank=LOCAL_RANK,
workers=workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '),
shuffle=True)
mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max()) # max label class
nb = len(train_loader) # number of batches
assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
# Process 0
if RANK in [-1, 0]:
val_loader = create_dataloader(val_path,
imgsz,
batch_size // WORLD_SIZE * 2,
gs,
single_cls,
hyp=hyp,
cache=None if noval else opt.cache,
rect=True,
rank=-1,
workers=workers * 2,
pad=0.5,
prefix=colorstr('val: '))[0]
if not resume:
labels = np.concatenate(dataset.labels, 0)
# c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
plot_labels(labels, names, save_dir)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
model.half().float() # pre-reduce anchor precision
callbacks.run('on_pretrain_routine_end')
# DDP mode
if cuda and RANK != -1:
model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)
# Model attributes
nl = de_parallel(
model).model[-1].nl # number of detection layers (to scale hyps)
hyp['box'] *= 3 / nl # scale to layers
hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers
hyp['obj'] *= (imgsz / 640)**2 * 3 / nl # scale to image size and layers
hyp['label_smoothing'] = opt.label_smoothing
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
last_opt_step = -1
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
stopper = EarlyStopping(patience=opt.patience)
compute_loss = ComputeLoss(model) # init loss class
LOGGER.info(
f'Image sizes {imgsz} train, {imgsz} val\n'
f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
f"Logging results to {colorstr('bold', save_dir)}\n"
f'Starting training for {epochs} epochs...')
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional, single-GPU only)
if opt.image_weights:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 / nc # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(range(dataset.n),
weights=iw,
k=dataset.n) # rand weighted idx
# Update mosaic border (optional)
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(3, device=device) # mean losses
if RANK != -1:
train_loader.sampler.set_epoch(epoch)
pbar = enumerate(train_loader)
LOGGER.info(
('\n' + '%10s' * 7) %
('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size'))
if RANK in [-1, 0]:
pbar = tqdm(
pbar, total=nb,
bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = nn.functional.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device)) # loss scaled by batch_size
if RANK != -1:
loss *= WORLD_SIZE # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize
if ni - last_opt_step >= accumulate:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
last_opt_step = ni
# Log
if RANK in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB)
pbar.set_description(('%10s' * 2 + '%10.4g' * 5) %
(f'{epoch}/{epochs - 1}', mem, *mloss,
targets.shape[0], imgs.shape[-1]))
callbacks.run('on_train_batch_end', ni, model, imgs, targets,
paths, plots, opt.sync_bn)
if callbacks.stop_training:
return
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for loggers
scheduler.step()
if RANK in [-1, 0]:
# mAP
callbacks.run('on_train_epoch_end', epoch=epoch)
ema.update_attr(model,
include=[
'yaml', 'nc', 'hyp', 'names', 'stride',
'class_weights'
])
final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
if not noval or final_epoch: # Calculate mAP
results, maps, _ = val.run(data_dict,
batch_size=batch_size //
WORLD_SIZE * 2,
imgsz=imgsz,
model=ema.ema,
single_cls=single_cls,
dataloader=val_loader,
save_dir=save_dir,
plots=False,
callbacks=callbacks,
compute_loss=compute_loss)
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
log_vals = list(mloss) + list(results) + lr
callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness,
fi)
# Save model
if (not nosave) or (final_epoch and not evolve): # if save
ckpt = {
'epoch': epoch,
'best_fitness': best_fitness,
'model': deepcopy(de_parallel(model)).half(),
'ema': deepcopy(ema.ema).half(),
'updates': ema.updates,
'optimizer': optimizer.state_dict(),
'wandb_id':
loggers.wandb.wandb_run.id if loggers.wandb else None,
'date': datetime.now().isoformat()
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
if (epoch > 0) and (opt.save_period >
0) and (epoch % opt.save_period == 0):
torch.save(ckpt, w / f'epoch{epoch}.pt')
del ckpt
callbacks.run('on_model_save', last, epoch, final_epoch,
best_fitness, fi)
# Stop Single-GPU
if RANK == -1 and stopper(epoch=epoch, fitness=fi):
break
# Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576
# stop = stopper(epoch=epoch, fitness=fi)
# if RANK == 0:
# dist.broadcast_object_list([stop], 0) # broadcast 'stop' to all ranks
# Stop DPP
# with torch_distributed_zero_first(RANK):
# if stop:
# break # must break all DDP ranks
# end epoch ----------------------------------------------------------------------------------------------------
# end training -----------------------------------------------------------------------------------------------------
if RANK in [-1, 0]:
LOGGER.info(
f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.'
)
for f in last, best:
if f.exists():
strip_optimizer(f) # strip optimizers
if f is best:
LOGGER.info(f'\nValidating {f}...')
results, _, _ = val.run(
data_dict,
batch_size=batch_size // WORLD_SIZE * 2,
imgsz=imgsz,
model=attempt_load(f, device).half(),
iou_thres=0.65 if is_coco else
0.60, # best pycocotools results at 0.65
single_cls=single_cls,
dataloader=val_loader,
save_dir=save_dir,
save_json=is_coco,
verbose=True,
plots=True,
callbacks=callbacks,
compute_loss=compute_loss) # val best model with plots
if is_coco:
callbacks.run('on_fit_epoch_end',
list(mloss) + list(results) + lr, epoch,
best_fitness, fi)
callbacks.run('on_train_end', last, best, plots, epoch, results)
LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
torch.cuda.empty_cache()
return results
class V5Centernet(nn.Module):
def __init__(self,
cfg,
num_classes=14,
pretrained=None,
device='cpu',
type='x'):
super().__init__()
self.model = Model(cfg, ch=3, nc=num_classes)
if pretrained:
# weights = torch.load(pretrained)
# self.model.load_state_dict(weights)
exclude = [] # exclude keys
ckpt = torch.load(pretrained,
map_location=device) # load checkpoint
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
self.model.state_dict(),
exclude=exclude) # intersect
self.model.load_state_dict(state_dict, strict=False) # load
print('Transferred %g/%g items from %s' %
(len(state_dict), len(
self.model.state_dict()), pretrained)) # report
del ckpt, state_dict
type = cfg.split('olov5')[-1][0]
if type == 'x':
channel_list = [1280, 640, 320, 160, 80]
elif type == 'l':
channel_list = [1024, 512, 256, 128, 64]
elif type == 'm':
channel_list = [768, 384, 192, 96, 48]
elif type == 's':
channel_list = [512, 256, 128, 64, 32]
else:
raise NotImplementedError(
f"model type {type} has not implemented!")
#upsampling's head
# self.center = nn.Sequential(
# nn.Conv2d(channel_list[0], 512, kernel_size=11, padding=5, bias=False),
# nn.BatchNorm2d(512),
# nn.ReLU(inplace=True),
# ).to(device)
# self.decode1 = ResDecode(channel_list[1] + 512, 256).to(device) #layer11 9
# self.decode2 = ResDecode(channel_list[2] + 256, 128).to(device) #layer8 6
# self.decode3 = ResDecode(channel_list[3] + 128, 64).to(device) #layer6 4
# self.decode4 = ResDecode(channel_list[4] + 64, 32).to(device) #layer3 2
# self.decode5 = ResDecode(32, 16).to(device) #layer2 0
# self.logit = nn.Conv2d(16, 1, kernel_size=3, padding=1) #segmentation output
self.mask = nn.Sequential(
nn.Conv2d(channel_list[0], 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 1, kernel_size=1, padding=0),
)
self.pooling = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Linear(channel_list[0], 1)
self.dropout = nn.Dropout(0.2)
def forward(self, x, augment=False):
y, dt = [], [] # outputs
bs = x.shape[0]
ipt = x.clone()
skip = []
for index, m in enumerate(self.model.model):
if m.f != -1: # if not from previous layer
x = y[m.f] if isinstance(
m.f, int) else [x if j == -1 else y[j]
for j in m.f] # from earlier layers
x = m(x) # run
y.append(x if m.i in self.model.save else None) # save output
# if not isinstance(x, list):
# print(index, x.shape)
if index in [0, 2, 4, 6, 9]:
skip.append(x)
if index == 9:
# z = self.center(skip[-1])
# # print('z shape: ', z.shape)
# z = self.decode1([skip[-2], resize_like(z, skip[-2])]) # ; print('d1',x.size())
# # print('z shape: ', z.shape)
# z = self.decode2([skip[-3], resize_like(z, skip[-3])]) # ; print('d2',x.size())
# # print('z shape: ', z.shape)
# z = self.decode3([skip[-4], resize_like(z, skip[-4])]) # ; print('d3',x.size())
# # print('z shape: ', z.shape)
# z = self.decode4([skip[-5], resize_like(z, skip[-5])]) # ; print('d4',x.size())
# # print('z shape: ', z.shape)
# z = self.decode5([resize_like(z, ipt)])
# # print('z shape: ', z.shape)
# seg_logit = self.logit(z)
seg_logit = self.mask(skip[-1])
# print('z shape: ', seg_logit.shape)
features = self.pooling(skip[-1]).view(bs, -1)
output = self.fc(self.dropout(features))
return x, seg_logit, output
def train(hyp, opt, device, tb_writer=None):
logger.info(
colorstr('hyperparameters: ') + ', '.join(f'{k}={v}'
for k, v in hyp.items()))
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve # create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict
is_coco = opt.data.endswith('coco.yaml')
# Logging- Doing this before checking the dataset. Might update data_dict
loggers = {'wandb': None} # loggers dict
if rank in [-1, 0]:
opt.hyp = hyp # add hyperparameters
run_id = torch.load(weights).get('wandb_id') if weights.endswith(
'.pt') and os.path.isfile(weights) else None
wandb_logger = WandbLogger(opt,
Path(opt.save_dir).stem, run_id, data_dict)
loggers['wandb'] = wandb_logger.wandb
data_dict = wandb_logger.data_dict
if wandb_logger.wandb:
weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp # WandbLogger might update weights, epochs if resuming
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
names = ['item'] if opt.single_cls and len(
data_dict['names']) != 1 else data_dict['names'] # class names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) # load checkpoint
model = Model(opt.cfg or ckpt['model'].yaml,
ch=3,
nc=nc,
anchors=hyp.get('anchors')).to(device) # create
exclude = [
'anchor'
] if (opt.cfg or hyp.get('anchors')) and not opt.resume else [
] # exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
model = Model(opt.cfg, ch=3, nc=nc,
anchors=hyp.get('anchors')).to(device) # create
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
train_path = data_dict['train']
test_path = data_dict['val']
# Freeze
freeze = [] # parameter names to freeze (full or partial)
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply decay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
if opt.linear_lr:
lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp[
'lrf'] # linear
else:
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# EMA
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
ema.updates = ckpt['updates']
# Results
if ckpt.get('training_results') is not None:
results_file.write_text(
ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
gs = max(int(model.stride.max()), 32) # grid size (max stride)
nl = model.model[
-1].nl # number of detection layers (used for scaling hyp['obj'])
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# Trainloader
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '))
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# Process 0
if rank in [-1, 0]:
testloader = create_dataloader(
test_path,
imgsz_test,
batch_size * 2,
gs,
opt, # testloader
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5,
prefix=colorstr('val: '))[0]
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
if plots:
plot_labels(labels, names, save_dir, loggers)
if tb_writer:
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
model.half().float() # pre-reduce anchor precision
# DDP mode
if cuda and rank != -1:
model = DDP(
model,
device_ids=[opt.local_rank],
output_device=opt.local_rank,
# nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698
find_unused_parameters=any(
isinstance(layer, nn.MultiheadAttention)
for layer in model.modules()))
# Model parameters
hyp['box'] *= 3. / nl # scale to layers
hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers
hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers
hyp['label_smoothing'] = opt.label_smoothing
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
# Start training
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
compute_loss = ComputeLoss(model) # init loss class
logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n'
f'Using {dataloader.num_workers} dataloader workers\n'
f'Logging results to {save_dir}\n'
f'Starting training for {epochs} epochs...')
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if opt.image_weights:
# Generate indices
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 / nc # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(4, device=device) # mean losses
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'total', 'labels', 'img_size'))
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_items = compute_loss(
pred, targets.to(device)) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
# Plot
if plots and ni < 3:
f = save_dir / f'train_batch{ni}.jpg' # filename
Thread(target=plot_images,
args=(imgs, targets, paths, f),
daemon=True).start()
# if tb_writer:
# tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(torch.jit.trace(model, imgs, strict=False), []) # add model graph
elif plots and ni == 10 and wandb_logger.wandb:
wandb_logger.log({
"Mosaics": [
wandb_logger.wandb.Image(str(x), caption=x.name)
for x in save_dir.glob('train*.jpg') if x.exists()
]
})
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
ema.update_attr(model,
include=[
'yaml', 'nc', 'hyp', 'gr', 'names', 'stride',
'class_weights'
])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
wandb_logger.current_epoch = epoch + 1
results, maps, times = test.test(data_dict,
batch_size=batch_size * 2,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
verbose=nc < 50
and final_epoch,
plots=plots and final_epoch,
wandb_logger=wandb_logger,
compute_loss=compute_loss,
is_coco=is_coco)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results +
'\n') # append metrics, val_loss
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Log
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
if tb_writer:
tb_writer.add_scalar(tag, x, epoch) # tensorboard
if wandb_logger.wandb:
wandb_logger.log({tag: x}) # W&B
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
wandb_logger.end_epoch(best_result=best_fitness == fi)
# Save model
if (not opt.nosave) or (final_epoch and not opt.evolve): # if save
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
results_file.read_text(),
'model':
deepcopy(
model.module if is_parallel(model) else model).half(),
'ema':
deepcopy(ema.ema).half(),
'updates':
ema.updates,
'optimizer':
optimizer.state_dict(),
'wandb_id':
wandb_logger.wandb_run.id if wandb_logger.wandb else None
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
if wandb_logger.wandb:
if ((epoch + 1) % opt.save_period == 0
and not final_epoch) and opt.save_period != -1:
wandb_logger.log_model(last.parent,
opt,
epoch,
fi,
best_model=best_fitness == fi)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Plots
if plots:
plot_results(save_dir=save_dir) # save as results.png
if wandb_logger.wandb:
files = [
'results.png', 'confusion_matrix.png',
*[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]
]
wandb_logger.log({
"Results": [
wandb_logger.wandb.Image(str(save_dir / f), caption=f)
for f in files if (save_dir / f).exists()
]
})
# Test best.pt
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
if opt.data.endswith('coco.yaml') and nc == 80: # if COCO
for m in (last,
best) if best.exists() else (last): # speed, mAP tests
results, _, _ = test.test(opt.data,
batch_size=batch_size * 2,
imgsz=imgsz_test,
conf_thres=0.001,
iou_thres=0.7,
model=attempt_load(m, device).half(),
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
save_json=True,
plots=False,
is_coco=is_coco)
# Strip optimizers
final = best if best.exists() else last # final model
for f in last, best:
if f.exists():
strip_optimizer(f) # strip optimizers
if opt.bucket:
os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload
if wandb_logger.wandb and not opt.evolve: # Log the stripped model
wandb_logger.wandb.log_artifact(
str(final),
type='model',
name='run_' + wandb_logger.wandb_run.id + '_model',
aliases=['last', 'best', 'stripped'])
wandb_logger.finish_run()
else:
dist.destroy_process_group()
torch.cuda.empty_cache()
return results
def train(hyp, opt, device, tb_writer=None):
logger.info(f'Hyperparameters {hyp}')
"""
获取记录训练日志的路径:
训练日志包括：权重、tensorboard文件、超参数hyp、设置的训练参数opt(也就是epochs,batch_size等),result.txt
result.txt包括: 占GPU内存、训练集的GIOU loss, objectness loss, classification loss, 总loss,
targets的数量, 输入图片分辨率, 准确率TP/(TP+FP),召回率TP/P ;
测试集的mAP50, [email protected]:0.95, GIOU loss, objectness loss, classification loss.
还会保存batch<3的ground truth
"""
# 如果设置进化算法则不会传入tb_writer(则为None)，设置一个evolve文件夹作为日志目录
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
# 设置生成文件的保存路径
wdir = log_dir / 'weights' # weights directory
os.makedirs(wdir, exist_ok=True)
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = str(log_dir / 'results.txt')
# 获取轮次、批次、总批次(涉及到分布式训练)、权重、进程序号(主要用于分布式训练)
epochs, batch_size, total_batch_size, weights, rank = \
opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
# Save run settings
# 保存hyp和opt
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
# 获取数据路径
cuda = device.type != 'cpu'
# 设置随机种子
# 需要在每一个进程设置相同的随机种子，以便所有模型权重都初始化为相同的值，即确保神经网络每次初始化都相同
init_seeds(2 + rank)
# 加载数据配置信息
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
# torch_distributed_zero_first同步所有进程
# check_dataset检查数据集，如果没找到数据集则下载数据集(仅适用于项目中自带的yaml文件数据集)
with torch_distributed_zero_first(rank):
check_dataset(data_dict) # check
# 获取训练集、测试集图片路径
train_path = data_dict['train']
test_path = data_dict['val']
# 获取类别数量和类别名字
# 如果设置了opt.single_cls则为一类
nc, names = (1, ['item']) if opt.single_cls else (
int(data_dict['nc']),
data_dict['names']) # 保存data.yaml中的number classes, names
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data) # check
# Model
# 判断weights字符串是否以'.pt'为结尾。若是，则说明本次训练需要预训练模型
pretrained = weights.endswith('.pt')
if pretrained:
# 加载模型，从google云盘中自动下载模型
# 但通常会下载失败，建议提前下载下来放进weights目录
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights,
map_location=device) # load checkpoint 导入权重文件
"""
这里模型创建，可通过opt.cfg，也可通过ckpt['model'].yaml
这里的区别在于是否是resume，resume时会将opt.cfg设为空，
则按照ckpt['model'].yaml创建模型；
这也影响着下面是否除去anchor的key(也就是不加载anchor)，
如果resume，则加载权重中保存的anchor来继续训练；
主要是预训练权重里面保存了默认coco数据集对应的anchor，
如果用户自定义了anchor，再加载预训练权重进行训练，会覆盖掉用户自定义的anchor；
所以这里主要是设定一个，如果加载预训练权重进行训练的话，就去除掉权重中的anchor，采用用户自定义的；
如果是resume的话，就是不去除anchor，就权重和anchor一起加载， 接着训练；
参考https://github.com/ultralytics/yolov5/issues/459
所以下面设置了intersect_dicts，该函数就是忽略掉exclude中的键对应的值
"""
'''
ckpt:
{'epoch': -1,
'best_fitness': array([ 0.49124]),
'training_results': None,
'model': Model(
...
)
'optimizer': None
}
'''
if hyp.get('anchors'): # 用户自定义的anchors优先级大于权重文件中自带的anchors
ckpt['model'].yaml['anchors'] = round(
hyp['anchors']) # force autoanchor
# 创建并初始化yolo模型
model = Model(opt.cfg or ckpt['model'].yaml, ch=3,
nc=nc).to(device) # create
'''
model =
Model(
(model): Sequential(
(0): Focus(...)
...
(24): Detect(...)
)
)
'''
# 如果opt.cfg存在，或重新设置了'anchors'，则将预训练权重文件中的'anchors'参数清除，使用用户自定义的‘anchors’信息
exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [
] # exclude keys
# state_dict变量存放训练过程中需要学习的权重和偏执系数，state_dict 是一个python的字典格式,以字典的格式存储,然后以字典的格式被加载,而且只加载key匹配的项
# 将ckpt中的‘model’中的”可训练“的每一层的参数建立映射关系（如 'conv1.weight'： 数值...）存在state_dict中
state_dict = ckpt['model'].float().state_dict() # to FP32
# 加载除了与exclude以外，所有与key匹配的项的参数 即将权重文件中的参数导入对应层中
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) # intersect
# 将最终模型参数导入yolo模型
model.load_state_dict(state_dict, strict=False) # load
logger.info(
'Transferred %g/%g items from %s' %
(len(state_dict), len(model.state_dict()), weights)) # report
else:
# 不进行预训练，则直接创建并初始化yolo模型
model = Model(opt.cfg, ch=3, nc=nc).to(device) # create
# Freeze
#freeze = ['', ] # parameter names to freeze (full or partial)
freeze = ['model.%s.' % x for x in range(10)
] # 冻结带有'model.0.'-'model.9.'的所有参数 即冻结0-9层的backbone
if any(freeze):
for k, v in model.named_parameters():
if any(x in k for x in freeze):
print('freezing %s' % k)
v.requires_grad = False
# Optimizer
"""
nbs人为模拟的batch_size;
就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64，
也就是模型梯度累积了64/16=4(accumulate)次之后
再更新一次模型，变相的扩大了batch_size
"""
nbs = 64 # nominal batch size
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
# 根据accumulate设置权重衰减系数
hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
# 将模型分成三组(w权重参数（非bn层）, bias, 其他所有参数)优化
for k, v in model.named_parameters(): # named_parameters：网络层的名字和参数的迭代器
'''
(0): Focus(
(conv): Conv(
(conv): Conv2d(12, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(80, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): Hardswish()
)
)
k: 网络层可训练参数的名字所属 如： model.0.conv.conv.weight 或 model.0.conv.bn.weight 或 model.0.conv.bn.bias （Focus层举例）
v： 对应网络层的具体参数 如：对应model.0.conv.conv.weight的 size为(80，12，3，3)的参数数据 即 卷积核的数量为80，深度为12，size为3×3
'''
v.requires_grad = True # 设置当前参数在训练时保留梯度信息
if '.bias' in k:
pg2.append(v) # biases (所有的偏置参数)
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay (非bn层的权重参数w)
else:
pg0.append(v) # all else （网络层的其他参数）
# 选用优化器，并设置pg0组的优化方式
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
# 设置权重参数weights（非bn层）的优化方式
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
# 设置偏置参数bias的优化方式
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# 设置学习率衰减，这里为余弦退火方式进行衰减
# 就是根据以下公式lf,epoch和超参数hyp['lrf']进行衰减
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[
'lrf']) + hyp['lrf'] # cosine 匿名余弦退火函数
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Resume
# 初始化开始训练的epoch和最好的结果
# best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, [email protected], [email protected]:0.95]再求和所得
# 根据best_fitness来保存best.pt
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
# 加载优化器与best_fitness
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# Results
# 加载训练结果result.txt
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
# Epochs
# 加载上次断点模型中训练的轮次，并在此基础上继续训练
start_epoch = ckpt['epoch'] + 1
# 如果使用断点重训的同时发现 start_epoch= 0，则说明上次训练正常结束，不存在断点
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weights, epochs)
shutil.copytree(wdir, wdir.parent /
f'weights_backup_epoch{start_epoch - 1}'
) # save previous weights
# 如果新设置epochs小于加载的epoch，则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, ckpt['epoch'], epochs))
epochs += ckpt['epoch'] # finetune additional epochs
del ckpt, state_dict
# Image sizes
# 获取模型总步长和模型输入图片分辨率
gs = int(max(model.stride)) # grid size (max stride)
# 检查输入图片分辨率确保能够整除总步长gs
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
# 分布式训练,参照:https://github.com/ultralytics/yolov5/issues/475
# DataParallel模式,仅支持单机多卡,不支持混合精度训练
# rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式
# 如果 当前运行设备为gpu 且 进程编号=-1 且gpu数量大于1时 才会进行分布式训练 ，将model对象放入DataParallel容器即可进行分布式训练
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
# 实现多GPU之间的BatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
logger.info('Using SyncBatchNorm()')
# Exponential moving average
'''
EMA ： YOLOv5优化策略之一
EMA + SGD可提高模型鲁棒性
为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建
'''
ema = ModelEMA(model) if rank in [-1, 0] else None
# DDP mode
# 如果rank不等于-1,则使用DistributedDataParallel模式
# local_rank为gpu编号,rank为进程,例如rank=3，local_rank=0 表示第 3 个进程内的第 1 块 GPU。
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank)
# Trainloader
# class dataloader 和 dataset .
dataloader, dataset = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers)
# 获取标签中最大的类别值，并于类别数作比较, 如果小于类别数则表示有问题
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
'''
dataloader和testloader不同之处在于：
1. testloader：没有数据增强，rect=True（大概是测试图片保留了原图的长宽比）
2. dataloader：数据增强，保留了矩形框训练。
'''
# Process 0
if rank in [-1, 0]:
# local_rank is set to -1. Because only the first process is expected to do evaluation.
# testloader
ema.updates = start_epoch * nb // accumulate # set EMA updates
# testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt,
# hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=True,
# rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # testloader
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf.to(device))
plot_labels(labels, save_dir=log_dir)
if tb_writer:
# tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384
tb_writer.add_histogram('classes', c, 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# Model parameters
# 根据自己数据集的类别数设置分类损失的系数
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
# 设置类别数，超参数
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
"""
设置giou的值在objectness loss中做标签的系数, 使用代码如下
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)
这里model.gr=1，也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签
"""
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
# 根据labels初始化图片采样权重（图像类别所占比例高的采样频率低）
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
# 获取类别的名字
model.names = names
# Start training
t0 = time.time()
# 获取warm-up训练的迭代次数
nw = max(round(hyp['warmup_epochs'] * nb),
1e3) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
# 初始化mAP和results
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls, angleloss)
"""
设置学习率衰减所进行到的轮次，
目的是打断训练后，--resume接着训练也能正常的衔接之前的训练进行学习率衰减
"""
scheduler.last_epoch = start_epoch - 1 # do not move
# 通过torch1.6自带的api设置混合精度训练
scaler = amp.GradScaler(enabled=cuda)
"""
打印训练和测试输入图片分辨率
加载图片时调用的cpu进程数
从哪个epoch开始训练
"""
logger.info(
'Image sizes %g train, %g test\nUsing %g dataloader workers\nLogging results to %s\n'
'Starting training for %g epochs...' %
(imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs))
# 训练
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
# model设置为训练模式，其中training属性表示BatchNorm与Dropout层在训练阶段和测试阶段中采取的策略不同，通过判断training值来决定前向传播策略
model.train()
# Update image weights (optional)
# 加载图片权重（可选）
if opt.image_weights:
# Generate indices
"""
如果设置进行图片采样策略，
则根据前面初始化的图片采样权重model.class_weights以及maps配合每张图片包含的类别数
通过random.choices生成图片索引indices从而进行采样
"""
if rank in [-1, 0]:
cw = model.class_weights.cpu().numpy() * (
1 - maps)**2 # class weights
iw = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=cw) # image weights
dataset.indices = random.choices(
range(dataset.n), weights=iw,
k=dataset.n) # rand weighted idx
# Broadcast if DDP
# 如果是DDP模式,则广播采样策略
if rank != -1:
indices = (torch.tensor(dataset.indices)
if rank == 0 else torch.zeros(dataset.n)).int()
dist.broadcast(indices, 0)
if rank != 0:
dataset.indices = indices.cpu().numpy()
# Update mosaic border
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
# 初始化训练时打印的平均损失信息
mloss = torch.zeros(5, device=device) # mean losses
if rank != -1:
# DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子，
# 每次epoch不同，随机种子就不同
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 9) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls',
'angle', 'total', 'targets', 'img_size'))
if rank in [-1, 0]:
# tqdm 创建进度条，方便训练时 信息的展示
pbar = tqdm(pbar, total=nb) # progress bar
optimizer.zero_grad()
for i, (
imgs, targets, paths, _
) in pbar: # batch ------------------------------------------------------------
'''
i: batch_index, 第i个batch
imgs : torch.Size([batch_size, 3, resized_height, resized_weight])
targets : torch.Size = (该batch中的目标数量, [该image属于该batch的第几个图片, class, xywh, θ])
paths : List['img1_path','img2_path',......,'img-1_path'] len(paths)=batch_size
shapes : size= batch_size, 不进行mosaic时进行矩形训练时才有值
'''
# ni计算迭代的次数iteration
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warmup
"""
warmup训练(前nw次迭代)
在前nw次迭代中，根据以下方式选取accumulate和学习率
"""
if ni <= nw:
xi = [0, nw] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / total_batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
"""
bias的学习率从0.1下降到基准学习率lr*lf(epoch)，
其他的参数学习率从0增加到lr*lf(epoch).
lf为上面设置的余弦退火的衰减函数
"""
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
# 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
# 采用上采样下采样函数interpolate完成imgs尺寸的转变，模式设置为双线性插值
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
# 前向传播
with amp.autocast(enabled=cuda):
'''
训练时返回x
x list: [small_forward, medium_forward, large_forward] eg:small_forward.size=( batch_size, 3种scale框, size1, size2, no)
'''
pred = model(imgs) # forward
# Loss
# 计算损失，包括分类损失，objectness损失，框的回归损失
# loss为总损失值，loss_items为一个元组(lbox, lobj, lcls, langle, loss)
loss, loss_items = compute_loss(
pred, targets.to(device), model,
csl_label_flag=True) # loss scaled by batch_size
if rank != -1:
# 平均不同gpu之间的梯度
loss *= opt.world_size # gradient averaged between devices in DDP mode
# Backward
scaler.scale(loss).backward()
# Optimize
# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数
if ni % accumulate == 0:
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# Print
if rank in [-1, 0]:
# mloss (lbox, lobj, lcls, langle, loss)
# 打印显存，进行的轮次，损失，target的数量和图片的size等信息
mloss = (mloss * i + loss_items) / (i + 1
) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 7) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
# 进度条显示以上信息
pbar.set_description(s)
# Plot
# 将前三次迭代batch的标签框在图片上画出来并保存
if ni < 3:
f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename
result = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer and result is not None:
tb_writer.add_image(
f, result, dataformats='HWC',
global_step=epoch) # 存储的格式为[H, W, C]
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
if ema:
# 更新EMA的属性
# 添加include的属性
ema.update_attr(
model,
include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])
final_epoch = epoch + 1 == epochs
# # 判断该epoch是否为最后一轮
# if not opt.notest or final_epoch: # Calculate mAP
# # 对测试集进行测试，计算mAP等指标
# # 测试时使用的是EMA模型
# results, maps, times = test.test(opt.data,
# batch_size=total_batch_size,
# imgsz=imgsz_test,
# model=ema.ema,
# single_cls=opt.single_cls,
# dataloader=testloader,
# save_dir=log_dir,
# plots=epoch == 0 or final_epoch) # plot first and last
# Write
# 将测试指标写入result.txt
with open(results_file, 'a') as f:
f.write(
s + '%10.4g' * 8 % results +
'\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp %s gs://%s/results/results%s.txt' %
(results_file, opt.bucket, opt.name))
# Tensorboard
# 添加指标，损失等信息到tensorboard显示
if tb_writer:
tags = [
'train/box_loss',
'train/obj_loss',
'train/cls_loss',
'train/angle_loss', # train loss
'metrics/precision',
'metrics/recall',
'metrics/mAP_0.5',
'metrics/mAP_0.5:0.95',
'val/box_loss',
'val/obj_loss',
'val/cls_loss',
'val/angle_loss', # val loss
'x/lr0',
'x/lr1',
'x/lr2'
] # params
for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
# 更新best_fitness
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, [email protected], [email protected]]
if fi > best_fitness:
best_fitness = fi
# Save model
"""
保存模型，还保存了epoch，results，optimizer等信息，
optimizer信息在最后一轮完成后不会进行保存 未完成训练则保存该信息
model保存的是EMA的模型
"""
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
f.read(),
'model':
ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
if rank in [-1, 0]:
# Strip optimizers
"""
模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；
并且对模型进行model.half(), 将Float32的模型->Float16，
可以减少模型大小，提高inference速度
"""
n = opt.name if opt.name.isnumeric() else ''
fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt'
for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
if str(f2).endswith('.pt'): # is *.pt
strip_optimizer(f2) # strip optimizer
# 上传结果到谷歌云盘
os.system(
'gsutil cp %s gs://%s/weights' %
(f2, opt.bucket)) if opt.bucket else None # upload
# Finish
# 可视化results.txt文件
if not opt.evolve:
plot_results(save_dir=log_dir) # save as results.png
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
# 释放显存
dist.destroy_process_group() if rank not in [-1, 0] else None
torch.cuda.empty_cache()
return results
def train(hyp):
epochs = opt.epochs # 300
batch_size = opt.batch_size # 64
weights = opt.weights # initial training weights
# Configure
init_seeds(1)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict
train_path = data_dict['train']
test_path = data_dict['val']
nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Create model
model = Model(opt.cfg).to(device)
assert model.md['nc'] == nc, '%s nc=%g classes but %s nc=%g classes' % (
opt.data, nc, opt.cfg, model.md['nc'])
# Image sizes
gs = int(max(model.stride)) # grid size (max stride)
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
if v.requires_grad:
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
optimizer = optim.Adam(pg0, lr=hyp['lr0']) if opt.adam else \
optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g conv.weight, %g other' %
(len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# Load Model
google_utils.attempt_download(weights)
start_epoch, best_fitness = 0, 0.0
if weights.endswith('.pt'): # pytorch format
ckpt = torch.load(weights, map_location=device) # load checkpoint
# load model
try:
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if model.state_dict()[k].shape == v.shape
} # to FP32, filter
model.load_state_dict(ckpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s." \
% (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
# load optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# load results
if ckpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(ckpt['training_results']) # write results.txt
start_epoch = ckpt['epoch'] + 1
del ckpt
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O1',
verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: ((
(1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.9 + 0.1 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # do not move
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count(
) > 1 and torch.distributed.is_available():
dist.init_process_group(
backend='nccl', # distributed backend
init_method='tcp://127.0.0.1:9999', # init method
world_size=1, # number of nodes
rank=0) # node rank
model = torch.nn.parallel.DistributedDataParallel(model)
# pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html
# Dataset
dataset = LoadImagesAndLabels(
train_path,
imgsz,
batch_size,
augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect, # rectangular training
cache_images=opt.cache_images,
single_cls=opt.single_cls)
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (
mlc, nc, opt.cfg)
# Dataloader
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
num_workers=nw,
shuffle=not opt.
rect, # Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
# Testloader
testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(
test_path,
imgsz_test,
batch_size,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls),
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
# Model parameters
hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(
device) # attach class weights
model.names = data_dict['names']
# Class frequency
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# cf = torch.bincount(c.long(), minlength=nc) + 1.
# model._initialize_biases(cf.to(device))
if tb_writer:
plot_labels(labels)
tb_writer.add_histogram('classes', c, 0)
# Check anchors
if not opt.noautoanchor:
check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
# Exponential moving average
ema = torch_utils.ModelEMA(model)
# Start training
t0 = time.time()
nb = len(dataloader) # number of batches
n_burn = max(3 * nb,
1e3) # burn-in iterations, max(3 epochs, 1k iterations)
maps = np.zeros(nc) # mAP per class
results = (
0, 0, 0, 0, 0, 0, 0
) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
print('Using %g dataloader workers' % nw)
print('Starting training for %g epochs...' % epochs)
# torch.autograd.set_detect_anomaly(True)
for epoch in range(
start_epoch, epochs
): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 -
maps)**2 # class weights
image_weights = labels_to_image_weights(dataset.labels,
nc=nc,
class_weights=w)
dataset.indices = random.choices(range(dataset.n),
weights=image_weights,
k=dataset.n) # rand weighted idx
mloss = torch.zeros(4, device=device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls',
'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
# Burn-in
if ni <= n_burn:
xi = [0, n_burn] # x interp
# model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(
1,
np.interp(ni, xi, [1, nbs / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(
ni, xi,
[0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi,
[0.9, hyp['momentum']])
# Multi-scale
if opt.multi_scale:
sz = random.randrange(imgsz * 0.5,
imgsz * 1.5 + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]
] # new shape (stretched to gs-multiple)
imgs = F.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets.to(device), model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_cached() /
1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1),
mem, *mloss, targets.shape[0],
imgs.shape[-1])
pbar.set_description(s)
# Plot
if ni < 3:
f = 'train_batch%g.jpg' % i # filename
res = plot_images(images=imgs,
targets=targets,
paths=paths,
fname=f)
if tb_writer:
tb_writer.add_image(f,
res,
dataformats='HWC',
global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Scheduler
scheduler.step()
# mAP
ema.update_attr(model)
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(
opt.data,
batch_size=batch_size,
imgsz=imgsz_test,
save_json=final_epoch
and opt.data.endswith(os.sep + 'coco.yaml'),
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.4g' * 7 % results +
'\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp results.txt gs://%s/results/results%s.txt' %
(opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = [
'train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5',
'metrics/F1', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss'
]
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(
1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
ckpt = {
'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model':
ema.ema.module if hasattr(model, 'module') else ema.ema,
'optimizer':
None if final_epoch else optimizer.state_dict()
}
# Save last, best and delete
torch.save(ckpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
n = opt.name
if len(n):
n = '_' + n if not n.isnumeric() else n
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'],
[flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (
f2, opt.bucket)) if opt.bucket and ispt else None # upload
if not opt.evolve:
plot_results() # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1,
(time.time() - t0) / 3600))
dist.destroy_process_group(
) if device.type != 'cpu' and torch.cuda.device_count() > 1 else None
torch.cuda.empty_cache()
return results
def train(hyp, opt, device):
save_dir, epochs, batch_size, total_batch_size, weights, rank = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank
do_semi = opt.do_semi
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True)
last = wdir / 'last.pt'
best = wdir / 'best.pt'
results_file = save_dir / 'results.txt'
# Save run settings
with open(save_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
plots = not opt.evolve #create plots
cuda = device.type != 'cpu'
init_seeds(2 + rank)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.SafeLoader)
nc = 1 if opt.single_cls else int(data_dict['nc']) #number of classes
names = ['item'] if opt.single_cls and len(
data_dict['names']) != 1 else data_dict['names']
assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (
len(names), nc, opt.data)
# Model
pretrained = weights.endswith('.pt')
if pretrained:
with torch_distributed_zero_first(rank):
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=device) #load checkpoint
model = Model(opt.cfg or ckpt['model'].yaml,
ch=3,
nc=nc,
anchors=hyp.get('anchors')).to(device) #create
exclude = [
'anchor'
] if (opt.cfg or hyp.get('anchors')) and not opt.resume else [
] #exclude keys
state_dict = ckpt['model'].float().state_dict() # to FP32
state_dict = intersect_dicts(state_dict,
model.state_dict(),
exclude=exclude) #intersect
model.load_state_dict(state_dict, strict=False) #load
else:
model = Model(opt.cfg, ch=3, nc=nc,
anchors=hyp.get('anchors')).to(device)
with torch_distributed_zero_first(rank):
check_dataset(data_dict) #check
train_path = data_dict['train']
test_path = data_dict['val']
# Optimizer
nbs = 64
accumulate = max(round(nbs / total_batch_size),
1) # accumulate loss before optimizing
hyp['weight_decay'] *= total_batch_size * accumulate / nbs
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm2d):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
pg1.append(v.weight) # apply dacay
if opt.adam:
optimizer = optim.Adam(pg0,
lr=hyp['lr0'],
betas=(hyp['momentum'],
0.999)) # adjust betal to momentum
else:
optimizer = optim.SGD(pg0,
lr=hyp['lr0'],
momentum=hyp['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': hyp['weight_decay']
}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
del pg0, pg1, pg2
if opt.linear_lr:
lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp[
'lrf'] # linear
else:
lf = one_cycle(1, hyp['lrf'], epochs)
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
# EMA
ema = ModelEMA(model) if rank in [-1, 0] else None
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# optimizer
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = ckpt['best_fitness']
# EMA
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
ema.updates = ckpt['updates']
# Results
if ckpt.get('training_results') is not None:
results_file.write_text(
ckpt['training_results']) # write results.txt
# Epochs
start_epoch = ckpt['epoch'] + 1
if opt.resume:
assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (
weight, epochs)
if epochs < start_epoch:
epochs += ckpt['epoch']
del ckpt, state_dict
# Image sizes
gs = max(int(model.stride.max()), 32) # grid size (max stride)
nl = model.model[
-1].nl # number of detection layer (used for scaling hyp['obj])
imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size
] # verify imgsz are gs-multiples
# DP mode
if cuda and rank == -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
# SyncBatchNorm
if opt.sync_bn and cuda and rank != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(
device)
# Trainloader
if do_semi:
dataloader, dataset, unlabeldataloader = create_dataloader(
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '),
do_semi=opt.do_semi)
else:
dataloader, dataset = create_dataloader(
train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
rank=rank,
world_size=opt.world_size,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '),
do_semi=opt.do_semi)
# Train teacher model
mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class
nb = len(dataloader) # number of batches
assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (
mlc, nc, opt.data, nc - 1)
# process 0
if rank in [-1, 0]:
testloader = create_dataloader(
test_path,
imgsz_test,
batch_size * 2,
gs,
opt, # testloader
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
rank=-1,
world_size=opt.world_size,
workers=opt.workers,
pad=0.5,
prefix=colorstr('val: '),
do_semi=False)[0]
if not opt.resume:
labels = np.concatenate(dataset.labels, 0)
c = torch.tensor(labels[:, 0]) # classes
# Anchors
if not opt.noautoanchor:
check_anchors(dataset,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
model.half().float() # pre-reduce anchor precision
# DDP mode
if cuda and rank != 1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=opt.local_rank,
find_unused_parameters=any(
isinstance(layer, nn.MultiheadAttention)
for layer in model.modules()))
# Model parameters
hyp['box'] *= 3. / nl # scale to layers
hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers
hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers
hyp['label_smoothing'] = opt.label_smoothing
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
model.class_weights = labels_to_class_weights(
dataset.labels, nc).to(device) * nc # attach class weights
model.names = names
# Train teacher model --> burn in
t0 = time.time()
nw = max(round(hyp['warmup_epochs'] * nb),
1000) # number of warmup iterations, max(3 epochs, 1k iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0
) # P, R, [email protected], [email protected], val_loss(box, obj, cls)
scheduler.last_epoch = start_epoch - 1 # do not move
scaler = amp.GradScaler(enabled=cuda)
compute_loss = ComputeLoss(model) # init loss class
burnin_epochs = epochs / 2
# burn in
for epoch in range(start_epoch,
burnin_epochs): # epoch-------------------------
model.train()
nb = len(dataloader)
mloss = torch.zeros(4, device=device) # mean loss
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
if rank in [-1, 0]:
pbar = tqdm(pbar, total=nb)
optimizer.zero_grad()
for i, (imgs, targets, paths, _) in pbar:
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float(
) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0
# Warm up
if ni <= [0, nw]:
xi = [0, nw]
accumulate = max(
1, np.interp(ni, xi, [1, nbs / total_batch_size].round()))
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [
hyp['warmup_bias_lr'] if j == 2 else 0.0,
x['initial_lr'] * lf(epoch)
])
if 'momentum' in x:
x['momentum'] = np.interp(
ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Forward
with amp.autocast(enabled=cuda):
pred = model(imgs) # forward
loss, loss_item = compute_loss(
pred, targets.to(device)) # loss scaled by batch_size
if rank != -1:
loss *= opt.world_size # gradient averaged between device in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize
if ni % accumulate == 0:
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
# print
if rank in [-1, 0]:
mloss = (mloss * i + loss_item) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9
if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 +
'%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem,
*mloss, targets.shape[0], imgs.shape[-1])
pbar.set_description(s)
scheduler.step()
# DDP process 0 or single-GPU
if rank in [-1, 0]:
# mAP
ema.update_attr(model,
include=[
'yaml', 'nc', 'hyp', 'gr', 'names', 'stride',
'class_weights'
])
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
results, maps, times = test.test(data_dict,
batch_size=batch_size * 2,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
verbose=nc < 50
and final_epoch,
plots=plots and final_epoch,
compute_loss=compute_loss)
fi = fitness(np.array(results).reshape(
1, -1)) # weighted combination of [P, R, mAP@50, [email protected]]
if fi > best_fitness:
best_fitness = fi
if (not opt.nosave) or (final_epoch and not opt.evolve): # if save
ckpt = {
'epoch':
epoch,
'best_fitness':
best_fitness,
'training_results':
results_file.read_text(),
'model':
deepcopy(model.module if is_parallel(model) else model).half(),
'ema':
deepcopy(ema.ema).half(),
'updates':
ema.updates,
'optimizer':
optimizer.state_dict()
}
if best_fitness == fi:
torch.save(ckpt, best)
del ckpt
# end epoch ----------------------------------------------------------------------------
# end warm up
# get persudo label
# STAC
# first apply weak augmentation on unlabeled dataset then use teacher net to predict the persudo labels
# Then apply strong augmentation on unlabeled dataset, use student net to get the logists and compute the unlabeled loss.
model.eval()
img = []
target = []
Path = []
imgsz = opt.img_size
for idx, batch in tqdm(enumerate(unlabeldataloader),
total=len(unlabeldataloader)):
imgs0, _, path, _ = batch # from uint8 to float16
with torch.no_grad():
pred = model(imgs0.to(device, non_blocking=True).float() /
255.0)[0]
gn = torch.tensor(imgs0.shape)[[3, 2, 3, 2]]
pred = non_max_suppression(pred,
opt.conf_thres,
opt.iou_thres,
classes=opt.classes,
agnostic=opt.agnostic_nms)
for index, pre in enumerate(pred):
predict_number = len(pre)
if predict_number == 0:
continue
Class = pre[:, 5].view(predict_number, 1).cpu()
XYWH = (xyxy2xywh(pre[:, :4])).cpu()
XYWH /= gn
pre = torch.cat((torch.zeros(predict_number, 1), Class, XYWH),
dim=1)
img.append(imgs0[index])
target.append(pre)
Path.append(path[index])
unlabeldataset = semiDataset(img, target, Path)
del img, targets, Path
model.train()
