def load_model(weights, device):
# Load model
# model = attempt_load(weights, map_location=device) # load FP32 model
with open('data/coco128.yaml') as f:
data_dict = yaml.load(f, Loader=yaml.FullLoader)
model = Model('models/yolov5s.yaml').to(device)
model.names = data_dict['names']
model = model.fuse().eval()
ckpt = torch.load(weights, map_location=device)
ckpt['model'].float().fuse().eval()
print({
k: (v.shape, model.state_dict()[k].shape)
for k, v in ckpt['model'].float().state_dict().items()
if model.state_dict()[k].shape != v.shape
})
print({
k: (v.shape, ckpt['model'].float().state_dict()[k].shape)
for k, v in model.state_dict().items()
if ckpt['model'].float().state_dict()[k].shape != v.shape
})
ckpt['model'] = {
k: v
for k, v in ckpt['model'].float().state_dict().items()
if model.state_dict()[k].shape == v.shape
}
model.load_state_dict(ckpt['model'], strict=False)
for k, m in model.named_modules():
m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatability
if isinstance(m, Conv) and isinstance(m.act, nn.Hardswish):
m.act = Hardswish()
if isinstance(m, Conv) and isinstance(m.act, nn.SiLU):
m.act = SiLU()
return model
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
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, 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_graph(
torch.jit.trace(model, imgs, strict=False),
[]) # add 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()
# 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
# 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, wandb=None):
logger.info(f'Hyperparameters {hyp}')
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
opt.logdir) / 'evolve' # logging directory
log_dir = Path("/project/train/log")
wdir = Path("/project/train/models") # weights directory
os.makedirs(wdir / 'last', exist_ok=True)
os.makedirs(wdir / 'final', exist_ok=True)
last = wdir / 'last' / 'last.pt'
best = wdir / 'final' / 'best.pt'
results_file = str(log_dir / 'log.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):
print(data_dict)
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:
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
# 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 # 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),
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))
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]:
if i % 500 == 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 / f'train_batch{ni}.jpg') # 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
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
log_imgs=opt.log_imgs)
# 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/log%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
# logger.info(results)
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
}
ckpt = {'model': ema.ema}
# 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'log{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 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(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 train(hyp, opt, tb_writer=None):
logger.info(
colorstr('hyperparameters: ') + ', '.join(f'{k}={v}'
for k, v in hyp.items()))
save_dir, epochs, batch_size, weights = Path(
opt.save_dir), opt.epochs, opt.batch_size, opt.weights
# Directories
wdir = save_dir / 'weights'
wdir.mkdir(parents=True, exist_ok=True) # make dir
last = wdir / 'last.pkl'
best = wdir / 'best.pkl'
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 = not opt.no_cuda
if cuda:
jt.flags.use_cuda = 1
init_seeds(1)
with open(opt.data) as f:
data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict
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
model = Model(opt.cfg, ch=3, nc=nc) # create
pretrained = weights.endswith('.pkl')
if pretrained:
model.load(weights) # load
# 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']}")
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_modules():
if hasattr(v, 'bias') and isinstance(v.bias, jt.Var):
pg2.append(v.bias) # biases
if isinstance(v, nn.BatchNorm):
pg0.append(v.weight) # no decay
elif hasattr(v, 'weight') and isinstance(v.weight, jt.Var):
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 = optim.LambdaLR(optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
loggers = {} # loggers dict
start_epoch, best_fitness = 0, 0.0
# 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
# EMA
ema = ModelEMA(model)
# Trainloader
dataloader = create_dataloader(train_path,
imgsz,
batch_size,
gs,
opt,
hyp=hyp,
augment=True,
cache=opt.cache_images,
rect=opt.rect,
workers=opt.workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '))
mlc = np.concatenate(dataloader.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)
ema.updates = start_epoch * nb // accumulate # set EMA updates
testloader = create_dataloader(
test_path,
imgsz_test,
batch_size,
gs,
opt, # testloader
hyp=hyp,
cache=opt.cache_images and not opt.notest,
rect=True,
workers=opt.workers,
pad=0.5,
prefix=colorstr('val: '))
labels = np.concatenate(dataloader.labels, 0)
c = jt.array(labels[:, 0]) # classes
# cf = torch.bincount(c.int(), minlength=nc) + 1. # frequency
# model._initialize_biases(cf)
if plots:
plot_labels(labels, save_dir, loggers)
if tb_writer:
tb_writer.add_histogram('classes', c.numpy(), 0)
# Anchors
if not opt.noautoanchor:
check_anchors(dataloader,
model=model,
thr=hyp['anchor_t'],
imgsz=imgsz)
# 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
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(
dataloader.labels, nc) * 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
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
cw = model.class_weights.numpy() * (1 -
maps)**2 / nc # class weights
iw = labels_to_image_weights(dataloader.labels,
nc=nc,
class_weights=cw) # image weights
dataloader.indices = random.choices(
range(dataloader.n), weights=iw,
k=dataloader.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 = jt.zeros((4, )) # mean losses
pbar = enumerate(dataloader)
logger.info(
('\n' + '%10s' * 7) %
('Epoch', 'box', 'obj', 'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(pbar, total=nb) # progress bar
for i, (
imgs, targets, paths, _
) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.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 / 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.interpolate(imgs,
size=ns,
mode='bilinear',
align_corners=False)
# Forward
pred = model(imgs) # forward
loss, loss_items = compute_loss(pred, targets,
model) # loss scaled by batch_size
if opt.quad:
loss *= 4.
# Optimize
optimizer.step(loss)
if ema:
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
s = ('%10s' + '%10.4g' * 6) % ('%g/%g' %
(epoch, epochs - 1), *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
# end batch ------------------------------------------------------------------------------------------------
# end epoch ----------------------------------------------------------------------------------------------------
# Scheduler
lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard
scheduler.step()
# 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(data=opt.data,
batch_size=batch_size,
imgsz=imgsz_test,
model=ema.ema,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
plots=plots and final_epoch)
# 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:
if hasattr(x, "numpy"):
x = x.numpy()
tb_writer.add_scalar(tag, x, epoch) # tensorboard
# 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:
# Save last, best and delete
jt.save(ema.ema.state_dict(), last)
if best_fitness == fi:
jt.save(ema.ema.state_dict(), best)
# end epoch ----------------------------------------------------------------------------------------------------
# end training
# Strip optimizers
final = best if best.exists() else last # final model
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
# Test best.pkl
logger.info('%g epochs completed in %.3f hours.\n' %
(epoch - start_epoch + 1, (time.time() - t0) / 3600))
best_model = Model(opt.cfg)
best_model.load(str(final))
best_model = best_model.fuse()
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=best_model,
single_cls=opt.single_cls,
dataloader=testloader,
save_dir=save_dir,
save_json=save_json,
plots=False)
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(
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, 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 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, 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()
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
# 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,
# '/media/ivan/share/core50_350_1f/test.txt',
imgsz_test,
total_batch_size,
gs,
hyp=hyp,
augment=False,
# cache=opt.cache_images and not opt.notest,
rect=False,
rank=-1,
world_size=opt.world_size,
workers=opt.workers)[0]
root = '/media/ivan/share/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)
extMem = externalMemory(size=200)
print(f'external memory file: {extMem.get_memory_file()}')
train_paths = [
'/media/ivan/share/demoset/train_4.txt',
'/media/ivan/share/demoset/train_2.txt'
]
valid_paths = [
'/media/ivan/share/demoset/valid.txt',
'/media/ivan/share/demoset/valid.txt'
]
# prepare_params(hyp, opt, device, tb_writer=None, wandb=None)
for core_batch in range(2):
print(f'------------CORE50 itertaion №:{core_batch}------------')
train_on_large_batch(core_batch,
train_paths[core_batch],
valid_paths[core_batch],
model,
device,
logger,
imgsz=imgsz,
imgsz_test=imgsz_test,
gs=gs,
opt=opt,
hyp=hyp,
nc=nc,
log_dir=log_dir,
tb_writer=tb_writer,
names=names,
optimizer=optimizer,
extMem=extMem,
scheduler=scheduler,
lf=lf,
best_fitness=best_fitness,
wandb_run=wandb_run)
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=model,
single_cls=opt.single_cls,
dataloader=all_test_dataloader,
save_dir=log_dir / 'images' / str(core_batch),
# 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
class Detector:
def __init__(self, load_ext_mem=False):
self.logger = logging.getLogger(__name__)
self.device = torch.device('cuda:0')
opt, hyp = get_opt_and_hyp()
self.logger.info(f'Hyperparameters {hyp}')
# self.log_dir = Path(opt.logdir) # logging directory
self.log_dir = increment_dir(Path(opt.logdir) / 'exp', opt.name) # runs/exp1
self.log_dir = Path(self.log_dir)
wdir = self.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(self.log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(hyp, f, sort_keys=False)
with open(self.log_dir / 'opt.yaml', 'w') as f:
yaml.dump(vars(opt), f, sort_keys=False)
# Configure
init_seeds(2)
# with open(opt.data) as f:
# data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict
# check_dataset(data_dict) # check
self.nc = 6
self.names = ['???'] * 6
# assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data) # check
# Model
weights = opt.weights
pretrained = weights.endswith('.pt')
if pretrained:
attempt_download(weights) # download if not found locally
ckpt = torch.load(weights, map_location=self.device) # load checkpoint
if hyp.get('anchors'):
ckpt['model'].yaml['anchors'] = round(hyp['anchors']) # force autoanchor
self.model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=self.nc).to(self.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, self.model.state_dict(), exclude=exclude) # intersect
self.model.load_state_dict(state_dict, strict=False) # load
self.logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(self.model.state_dict()), weights)) # report
else:
self.model = Model(opt.cfg, ch=3, nc=self.nc).to(self.device) # create
# Freeze
# freeze = [] # parameter names to freeze (full or partial)
# for k, v in self.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 / opt.total_batch_size), 1) # accumulate loss before optimizing
hyp['weight_decay'] *= opt.total_batch_size * accumulate / nbs # scale weight_decay
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in self.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
self.optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
self.optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
self.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 / opt.epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine
#self.scheduler = lr_scheduler.LambdaLR(self.optimizer, lr_lambda=lf)
# plot_lr_scheduler(optimizer, scheduler, epochs)
# Resume
# start_epoch = 0
self.best_fitness = 0.0
if pretrained:
# Optimizer
if ckpt['optimizer'] is not None:
self.optimizer.load_state_dict(ckpt['optimizer'])
self.best_fitness = ckpt['best_fitness']
# Epochs
# start_epoch = ckpt['epoch'] + 1
del ckpt, state_dict
# Image sizes
self.gs = int(max(self.model.stride)) # grid size (max stride)
# Model parameters
hyp['cls'] *= self.nc / 80. # scale coco-tuned hyp['cls'] to current dataset
self.model.nc = self.nc # attach number of classes to model
self.model.hyp = hyp # attach hyperparameters to model
self.model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou)
# extenral memory
self.extMem = externalMemory(size=200)
print(f'external memory file: {self.extMem.get_memory_file()}')
self.opt = opt
self.hyp = hyp
self.added_classes = 0
# self.train.n_iter = 0
def train(self, train_files, class_names, valid_file=None):
# self.log_dir = Path(str(self.log_dir) + str(train.n_iter))
if not isinstance(class_names, list):
class_names = [class_names]
n_classes_to_add = len(class_names)
if not isinstance(train_files, list):
train_files = [train_files]
fix_class_id(train_files, class_names, self.added_classes)
# fix_class_id([valid_file], class_names, self.added_classes)
use_ext_mem = False if self.added_classes == 0 else True
for cn in class_names:
self.names[self.added_classes] = cn
self.added_classes += 1
train_on_large_batch(n_classes_to_add, train_files, self.model, self.device, self.logger,
valid_path=valid_file, imgsz=self.opt.img_size,
imgsz_test=self.opt.img_size, gs=self.gs,
opt=self.opt, hyp=self.hyp, nc=self.nc, log_dir=self.log_dir, tb_writer=None,
names=self.names, optimizer=self.optimizer,
extMem=self.extMem, best_fitness=self.best_fitness, use_ext_mem=use_ext_mem)
#self.model = attempt_load('runs/weights/best.pt')
def save_state(self, path='detector.pckl'):
with open(path, 'wb') as f:
pickle.dump(self, f)
def detect(self, image):
# image: array
# return [new_image, bboxes]
return detect_img(image, self.model, augment=True)
def find_object(self, object_name:str, image):
# returns x,y,w,h of found object
# if unknown object return "unknown object"
# if not found return "not found"
det = detect_img(image, self.model)[1]
if object_name not in self.names:
print(f'Unknown object:{object_name} in list: {self.names}')
return
obj_index = self.names.index(object_name)
ret_bboxes = det[det[..., -1] == obj_index]
if ret_bboxes.nelement() == 0:
print(f'{object_name} not found')
return
return ret_bboxes # each bbox: [x1, y1, x2, y2, confidence, class_index]
