def train(config, ADMM, device, train_loader, criterion, optimizer, scheduler,
epoch):
config.model.train()
ce_loss = None
for batch_idx, (data, target) in enumerate(train_loader):
# adjust learning rate
if config.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, config)
else:
if scheduler is not None:
scheduler.step()
data, target = data.to(device), target.to(device)
if config.gpu is not None:
data = data.cuda(config.gpu, non_blocking=True)
target = target.cuda(config.gpu, non_blocking=True)
if config.mixup:
data, target_a, target_b, lam = mixup_data(data, target,
config.alpha)
optimizer.zero_grad()
output = config.model(data)
if config.mixup:
ce_loss = mixup_criterion(criterion, output, target_a, target_b,
lam, config.smooth)
else:
ce_loss = criterion(output, target, smooth=config.smooth)
if config.admm:
admm.admm_update(config, ADMM, device, train_loader, optimizer,
epoch, data, batch_idx) # update Z and U
ce_loss, admm_loss, mixed_loss = admm.append_admm_loss(
config, ADMM, ce_loss) # append admm losss
if config.admm:
mixed_loss.backward()
else:
ce_loss.backward()
if config.masked_progressive:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.zero_masks:
W.grad *= config.zero_masks[name]
if config.masked_retrain:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.masks:
W.grad *= config.masks[name]
optimizer.step()
if batch_idx % config.print_freq == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), ce_loss.item()))
def train(train_loader, config, ADMM, criterion, optimizer, scheduler, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to train mode
config.model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
# adjust learning rate
if config.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, config)
else:
scheduler.step()
input = input.cuda(config.gpu, non_blocking=True)
target = target.cuda(config.gpu)
data = input
if config.mixup:
input, target_a, target_b, lam = mixup_data(
input, target, config.alpha)
# compute output
output = config.model(input)
if config.mixup:
ce_loss = mixup_criterion(criterion, output, target_a, target_b,
lam, config.smooth)
else:
ce_loss = criterion(output, target, smooth=config.smooth)
if config.admm:
admm.admm_update(config, ADMM, device, train_loader, optimizer,
epoch, data, i) # update Z and U
ce_loss, admm_loss, mixed_loss = admm.append_admm_loss(
config, ADMM, ce_loss) # append admm losss
# measure accuracy and record loss
acc1, acc5 = accuracy(output, target, topk=(1, 5))
losses.update(ce_loss.item(), input.size(0))
top1.update(acc1[0], input.size(0))
top5.update(acc5[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
if config.admm:
mixed_loss.backward()
else:
ce_loss.backward()
if config.masked_progressive:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.zero_masks:
W.grad *= config.zero_masks[name]
if config.masked_retrain:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.masks:
W.grad *= config.masks[name]
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % config.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Acc@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1,
top5=top5))
print("cross_entropy loss: {}".format(ce_loss))
def train(hyp):
# batch_time = AverageMeter()
# data_time = AverageMeter()
# losses = AverageMeter()
cfg = opt.cfg
data = opt.data
epochs = opt.epochs # 500200 batches at bs 64, 117263 images = 273 epochs
batch_size = opt.batch_size
accumulate = max(round(64 / batch_size), 1) # accumulate n times before optimizer update (bs 64)
weights = opt.weights # initial training weights
imgsz_min, imgsz_max, imgsz_test = opt.img_size # img sizes (min, max, test)
# Image Sizes
gs = 32 # (pixels) grid size
assert math.fmod(imgsz_min, gs) == 0, '--img-size %g must be a %g-multiple' % (imgsz_min, gs)
opt.multi_scale |= imgsz_min != imgsz_max # multi if different (min, max)
if opt.multi_scale:
if imgsz_min == imgsz_max:
imgsz_min //= 1.5
imgsz_max //= 0.667
grid_min, grid_max = imgsz_min // gs, imgsz_max // gs
imgsz_min, imgsz_max = int(grid_min * gs), int(grid_max * gs)
img_size = imgsz_max # initialize with max size
# Configure run
init_seeds()
data_dict = parse_data_cfg(data)
train_path = data_dict['train']
test_path = data_dict['valid']
nc = 1 if opt.single_cls else int(data_dict['classes']) # number of classes
hyp['cls'] *= nc / 80 # update coco-tuned hyp['cls'] to current dataset
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Initialize model
model = Darknet(cfg).to(device)
# Optimizer
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in dict(model.named_parameters()).items():
if '.bias' in k:
pg2 += [v] # biases
elif 'Conv2d.weight' in k:
pg1 += [v] # apply weight_decay
else:
pg0 += [v] # all else
if opt.adam:
# hyp['lr0'] *= 0.1 # reduce lr (i.e. SGD=5E-3, Adam=5E-4)
optimizer = optim.Adam(pg0, lr=hyp['lr0'])
# optimizer = AdaBound(pg0, lr=hyp['lr0'], final_lr=0.1)
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g Conv2d.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
start_epoch = 0
best_fitness = 0.0
# attempt_download(weights)
if opt.freeze_layers:
output_layer_indices = [idx - 1 for idx, module in enumerate(model.module_list) if isinstance(module, YOLOLayer)]
freeze_layer_indices = [x for x in range(len(model.module_list)) if
(x not in output_layer_indices) and
(x - 1 not in output_layer_indices)]
for idx in freeze_layer_indices:
for parameter in model.module_list[idx].parameters():
parameter.requires_grad_(False)
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.95 + 0.05 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # see link below
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# Plot lr schedule
# y = []
# for _ in range(epochs):
# scheduler.step()
# y.append(optimizer.param_groups[0]['lr'])
# plt.plot(y, '.-', label='LambdaLR')
# plt.xlabel('epoch')
# plt.ylabel('LR')
# plt.tight_layout()
# plt.savefig('LR.png', dpi=300)
# Dataset
dataset = LoadImagesAndLabels(train_path, img_size, batch_size,
augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect, # rectangular training
cache_images=opt.cache_images,
single_cls=opt.single_cls)
# Dataloader
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=nw,
shuffle=not opt.rect,
# Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
# Testloader
testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls),
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
initial_rho = opt.rho
t0 = time.time()
"""====================="""
""" multi-rho admm train"""
"""====================="""
if opt.admm:
opt.notest = True
# possible weights are '*.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc.
chkpt = torch.load(weights, map_location=device)
# load model
try:
# chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt['model'], strict=False)
except Exception as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
"See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
print(e)
raise KeyError(s) from e
del chkpt
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
dist.init_process_group(backend='nccl', # 'distributed backend'
init_method='tcp://127.0.0.1:9999', # distributed training init method
world_size=1, # number of nodes for distributed training
rank=0) # distributed training node rank
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level
# Model parameters
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
# Model EMA
ema = torch_utils.ModelEMA(model)
# Start training
nb = len(dataloader) # number of batches
n_burn = max(int(0.7 * nb), 500) # burn-in iterations, max(0.7 epochs, 500 iterations)
maps = np.zeros(nc) # mAP per class
# torch.autograd.set_detect_anomaly(True)
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
print('Image sizes %g - %g train, %g test' % (imgsz_min, imgsz_max, imgsz_test))
print('Using %g dataloader workers' % nw)
print('Starting training for %g epochs...' % epochs)
for i in range(opt.rho_num):
current_rho = initial_rho * 10 ** i
ADMM = admm.ADMM(model, file_name="./prune_config/" + opt.config_file + ".yaml", rho=current_rho)
admm.admm_initialization(opt, ADMM=ADMM, model=model) # intialize Z variable
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
print("current rho: {}".format(current_rho))
model.train()
masks = {}
if opt.masked_retrain and not opt.combine_progressive:
print("full acc re-train masking")
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
if name not in ADMM.prune_ratios:
continue
above_threshold, W = admm.weight_pruning(opt, W, ADMM.prune_ratios[name])
W.data = W
masks[name] = above_threshold
elif opt.combine_progressive:
print("progressive admm-train/re-train masking")
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
weight = W.cpu().detach().numpy()
non_zeros = weight != 0
non_zeros = non_zeros.astype(np.float32)
zero_mask = torch.from_numpy(non_zeros).cuda()
W = torch.from_numpy(weight).cuda()
W.data = W
masks[name] = zero_mask
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx
mloss = torch.zeros(4).to(device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device).float() / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
# Burn-in
if ni <= n_burn:
xi = [0, n_burn] # x interp
model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(1, np.interp(ni, xi, [1, 64 / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
x['weight_decay'] = np.interp(ni, xi, [0.0, hyp['weight_decay'] if j == 1 else 0.0])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])
# Multi-Scale
if opt.multi_scale:
if ni / accumulate % 1 == 0: # adjust img_size (67% - 150%) every 1 batch
img_size = random.randrange(grid_min, grid_max + 1) * gs
sf = img_size / 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 32-multiple)
imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets, model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
loss *= batch_size / 64 # scale loss
admm.z_u_update(opt, ADMM, model, device, dataloader, optimizer, epoch, imgs, i,
tb_writer) # update Z and U variables
loss, admm_loss, mixed_loss = admm.append_admm_loss(opt, ADMM, model,
loss) # append admm losss
if mixed_precision:
with amp.scale_loss(mixed_loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
mixed_loss.backward()
if opt.combine_progressive:
with torch.no_grad():
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
if name in masks:
W.grad *= masks[name]
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.3g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, len(targets), img_size)
pbar.set_description(s)
# Plot
# if ni < 1:
# f = 'train_batch%g.jpg' % i # filename
# res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
# if tb_writer:
# tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
# # tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Update scheduler
if opt.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, opt)
else:
scheduler.step()
# Process epoch results
ema.update_attr(model)
final_epoch = epoch + 1 == epochs
if not opt.notest: # Calculate mAP #or final_epoch
is_coco = any([x in data for x in ['coco.data', 'coco2014.data', 'coco2017.data']]) and model.nc == 80
results, maps = test.test(cfg,
data,
batch_size=batch_size,
imgsz=imgsz_test,
model=ema.ema,
save_json=final_epoch and is_coco,
single_cls=opt.single_cls,
dataloader=testloader,
multi_label=ni > n_burn)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1',
'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# end epoch ----------------------------------------------------------------------------------------------------
# end training
# admm_adjust_learning_rate ----------------------------------------------------------------------------------------------------
admm.admm_adjust_learning_rate(optimizer, epoch, opt)
# end admm_adjust_learning_rate ----------------------------------------------------------------------------------------------------
print("Saving model.")
torch.save(
model.module.state_dict() if type(model) is nn.parallel.DistributedDataParallel else model.state_dict(),
"./model_pruned/yolov4_{}_{}_{}.pt".format(
current_rho, opt.config_file, opt.sparsity_type))
if not opt.evolve:
plot_results() # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
# dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
# torch.cuda.empty_cache()
# return results
"""=============="""
"""masked retrain"""
"""=============="""
if opt.masked_retrain:
ADMM = admm.ADMM(model, file_name="./prune_config/" + opt.config_file + ".yaml", rho=initial_rho)
if not opt.resume:
# possible weights are '*.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc.
print("\n>_ Loading file: ./model_pruned/yolov4_{}_{}_{}.pt".format(initial_rho * 10 ** (opt.rho_num - 1), opt.config_file, opt.sparsity_type))
chkpt = torch.load("./model_pruned/yolov4_{}_{}_{}.pt".format(initial_rho * 10 ** (opt.rho_num - 1), opt.config_file, opt.sparsity_type), map_location=device)
# chkpt = torch.load(weights, map_location=device)
# load model
try:
# chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt, strict=False) #['model']
except KeyError as e:
# s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
# "See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
raise KeyError() from e
#----------------------------------------------hard prune------------------------------------------------
admm.hard_prune(opt, ADMM, model)
#----------------------------------------------hard prune------------------------------------------------
else:
try:
chkpt = torch.load(weights, map_location=device)
chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt['model'], strict=False)
except KeyError as e:
# s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
# "See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
raise KeyError() from e
# load optimizer
if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer'])
best_fitness = chkpt['best_fitness']
# load results
if chkpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(chkpt['training_results']) # write results.txt
start_epoch = chkpt['epoch'] + 1
del chkpt
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
dist.init_process_group(backend='nccl', # 'distributed backend'
init_method='tcp://127.0.0.1:9999', # distributed training init method
world_size=1, # number of nodes for distributed training
rank=0) # distributed training node rank
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level
# Model parameters
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
# Model EMA
ema = torch_utils.ModelEMA(model)
# Start training
nb = len(dataloader) # number of batches
n_burn = max(3 * nb, 500) # burn-in iterations, max(3 epochs, 500 iterations)
maps = np.zeros(nc) # mAP per class
# torch.autograd.set_detect_anomaly(True)
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
print('Image sizes %g - %g train, %g test' % (imgsz_min, imgsz_max, imgsz_test))
print('Using %g dataloader workers' % nw)
print('Starting training for %g epochs...' % epochs)
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
model.train()
if opt.masked_retrain and not opt.combine_progressive:
print("full acc re-train masking")
masks = {}
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
if name not in ADMM.prune_ratios:
continue
above_threshold, W = admm.weight_pruning(opt, W, ADMM.prune_ratios[name])
W.data = W
masks[name] = above_threshold
elif opt.combine_progressive:
print("progressive admm-train/re-train masking")
masks = {}
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
weight = W.cpu().detach().numpy()
non_zeros = weight != 0
non_zeros = non_zeros.astype(np.float32)
zero_mask = torch.from_numpy(non_zeros).cuda()
W = torch.from_numpy(weight).cuda()
W.data = W
masks[name] = zero_mask
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights,
k=dataset.n) # rand weighted idx
mloss = torch.zeros(4).to(device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device).float() / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
# Burn-in
if ni <= n_burn:
xi = [0, n_burn] # x interp
model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(1, np.interp(ni, xi, [1, 64 / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
x['weight_decay'] = np.interp(ni, xi, [0.0, hyp['weight_decay'] if j == 1 else 0.0])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])
# Multi-Scale
if opt.multi_scale:
if ni / accumulate % 1 == 0: # adjust img_size (67% - 150%) every 1 batch
img_size = random.randrange(grid_min, grid_max + 1) * gs
sf = img_size / 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 32-multiple)
imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
pred = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets, model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
loss *= batch_size / 64 # scale loss
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if opt.combine_progressive:
with torch.no_grad():
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
if name in masks:
W.grad *= masks[name]
if opt.masked_retrain:
with torch.no_grad():
for name, W in (model.module.named_parameters() if type(
model) is torch.nn.parallel.DistributedDataParallel else model.named_parameters()):
if name in masks:
W.grad *= masks[name]
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.3g' * 6) % (
'%g/%g' % (epoch, epochs - 1), mem, *mloss, len(targets), img_size)
pbar.set_description(s)
# Plot
if ni < 1:
f = 'train_batch%g.jpg' % i # filename
res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
if tb_writer:
tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Update scheduler
scheduler.step()
# Process epoch results
ema.update_attr(model)
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
is_coco = any(
[x in data for x in ['coco.data', 'coco2014.data', 'coco2017.data']]) and model.nc == 80
results, maps = test.test(cfg,
data,
batch_size=batch_size,
imgsz=imgsz_test,
model=ema.ema,
save_json=final_epoch and is_coco,
single_cls=opt.single_cls,
dataloader=testloader,
multi_label=ni > n_burn)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1',
'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness: #results[2]
best_fitness = fi #results[2]
print("\n>_ Got better accuracy {:.3f}% now...\n".format(results[2]))
# torch.save(ema.ema.module.state_dict() if hasattr(model, 'module') else ema.ema.state_dict(),
# "./model_retrained/yolov4_retrained_acc_{:.3f}_{}rhos_{}_{}.pt".format(results[2], opt.rho_num, opt.config_file, opt.sparsity_type))
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if save:
with open(results_file, 'r') as f: # create checkpoint
chkpt = {'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': ema.ema.module.state_dict() if hasattr(model,
'module') else ema.ema.state_dict(),
'optimizer': None if final_epoch else optimizer.state_dict()}
# Save last, best and delete
torch.save(chkpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(chkpt, best)
del chkpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
test_sparsity(model)
print("Best Acc: {:.4f}".format(results[2]))
n = opt.name
if len(n):
n = '_' + n if not n.isnumeric() else n
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (
f2, opt.bucket)) if opt.bucket and ispt else None # upload
if not opt.evolve:
plot_results() # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
# dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
# torch.cuda.empty_cache()
return results
def train(train_loader, criterion, optimizer, epoch, config):
batch_time = AverageMeter()
data_time = AverageMeter()
nat_losses = AverageMeter()
adv_losses = AverageMeter()
nat_loss = 0
adv_loss = 0
nat_top1 = AverageMeter()
adv_top1 = AverageMeter()
# switch to train mode
config.model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
# adjust learning rate
if config.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, config)
else:
scheduler.step()
if config.gpu is not None:
input = input.cuda(config.gpu, non_blocking=True)
target = target.cuda(config.gpu, non_blocking=True)
if config.mixup:
input, target_a, target_b, lam = mixup_data(
input, target, config.alpha)
# compute output
nat_output, adv_output, pert_inputs = config.model(input, target)
if config.mixup:
adv_loss = mixup_criterion(criterion, adv_output, target_a,
target_b, lam, config.smooth)
nat_loss = mixup_criterion(criterion, nat_output, target_a,
target_b, lam, config.smooth)
else:
adv_loss = criterion(adv_output, target, smooth=config.smooth)
nat_loss = criterion(nat_output, target, smooth=config.smooth)
if config.admm:
admm.admm_update(config, ADMM, device, train_loader, optimizer,
epoch, input, i) # update Z and U
adv_loss, admm_loss, mixed_loss = admm.append_admm_loss(
config, ADMM, adv_loss) # append admm losss
# measure accuracy and record loss
nat_acc1, _ = accuracy(nat_output, target, topk=(1, 5))
adv_acc1, _ = accuracy(adv_output, target, topk=(1, 5))
nat_losses.update(nat_loss.item(), input.size(0))
adv_losses.update(adv_loss.item(), input.size(0))
adv_top1.update(adv_acc1[0], input.size(0))
nat_top1.update(nat_acc1[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
if config.admm:
mixed_loss.backward()
else:
adv_loss.backward()
if config.masked_progressive:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.zero_masks:
W.grad *= config.zero_masks[name]
if config.masked_retrain:
with torch.no_grad():
for name, W in config.model.named_parameters():
if name in config.masks:
W.grad *= config.masks[
name] #returns boolean array called mask when weights are above treshhold
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % config.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Nat_Loss {nat_loss.val:.4f} ({nat_loss.avg:.4f})\t'
'Nat_Acc@1 {nat_top1.val:.3f} ({nat_top1.avg:.3f})\t'
'Adv_Loss {adv_loss.val:.4f} ({adv_loss.avg:.4f})\t'
'Adv_Acc@1 {adv_top1.val:.3f} ({adv_top1.avg:.3f})\t'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
nat_loss=nat_losses,
nat_top1=nat_top1,
adv_loss=adv_losses,
adv_top1=adv_top1))
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--config_file', type=str, default='', help ="config file")
parser.add_argument('--stage', type=str, default='', help ="select the pruning stage")
args = parser.parse_args()
config = Config(args)
use_cuda = True
init = Init_Func(config.init_func)
torch.manual_seed(config.random_seed)
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor()
#transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=64, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor()
#transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=1000, shuffle=True, **kwargs)
model = None
if config.arch == 'lenet_bn':
model = LeNet_BN().to(device)
elif config.arch == 'lenet':
model = LeNet().to(device)
elif config.arch == 'lenet_adv':
model = LeNet_adv(w=config.width_multiplier).to(device)
if config.arch not in model_names:
raise Exception("unknown model architecture")
### for initialization experiments
for name,W in model.named_parameters():
if 'conv' in name and 'bias' not in name:
print ('initialization uniform')
#W.data = torch.nn.init.uniform_(W.data)
W.data = init.init(W.data)
model = AttackPGD(model,config)
#### loading initialization
'''
### for lottery tickets experiments
read_dict = np.load('lenet_adv_retrained_w16_1_cut.pt_init.npy').item()
for name,W in model.named_parameters():
if name not in read_dict:
continue
print (name)
#print ('{} has shape {}'.format(name,read_dict[name].shape))
print (read_dict[name].shape)
W.data = torch.from_numpy(read_dict[name])
'''
config.model = model
if config.load_model:
# unlike resume, load model does not care optimizer status or start_epoch
print('==> Loading from {}'.format(config.load_model))
config.model.load_state_dict(torch.load(config.load_model, map_location=lambda storage, loc: storage))
#config.model.load_state_dict(torch.load(config.load_model,map_location = {'cuda:0':'cuda:{}'.format(config.gpu)}))
torch.cuda.set_device(config.gpu)
config.model.cuda(config.gpu)
test(config, device, test_loader)
ADMM = None
config.prepare_pruning()
if config.admm:
ADMM = admm.ADMM(config)
optimizer = None
if (config.optimizer == 'sgd'):
optimizer = torch.optim.SGD(config.model.parameters(), config.lr,
momentum=0.9,
weight_decay=1e-6)
elif (config.optimizer =='adam'):
optimizer = torch.optim.Adam(config.model.parameters(),config.lr)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=config.epochs*len(train_loader),eta_min=4e-08)
if config.resume:
if os.path.isfile(config.resume):
checkpoint = torch.load(config.resume)
config.start_epoch = checkpoint['epoch']
best_adv_acc = checkpoint['best_adv_acc']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(config.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(config.resume))
if config.masked_retrain:
# make sure small weights are pruned and confirm the acc
print ("<============masking both weights and gradients for retrain")
admm.masking(config)
print ("<============testing sparsity before retrain")
admm.test_sparsity(config)
test(config, device, test_loader)
if config.masked_progressive:
admm.zero_masking(config)
for epoch in range(0, config.epochs+1):
if config.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, config)
else:
if config.lr_scheduler == 'cosine':
scheduler.step()
elif config.lr_scheduler == 'sgd':
if epoch == 20:
config.lr/=10
for param_group in optimizer.param_groups:
param_group['lr'] = config.lr
else:
pass # it uses adam
train(config,ADMM,device, train_loader, optimizer, epoch)
test(config, device, test_loader)
admm.test_sparsity(config)
test(config, device, test_loader)
if config.save_model and config.admm:
print ('saving model {}'.format(config.save_model))
torch.save(config.model.state_dict(),config.save_model)
def prune_train(args, pre_mask, ADMM, train_loader, criterion, optimizer,
scheduler, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
idx_loss_dict = {}
# switch to train mode
model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
target = target.long().cuda()
# measure data loading time
data_time.update(time.time() - end)
# adjust learning rate
if args.admm:
admm.admm_adjust_learning_rate(optimizer, epoch, args)
else:
scheduler.step()
input = input.float().cuda()
if args.mixup:
input, target_a, target_b, lam = mixup_data(
input, target, args.alpha)
# compute output
output = model(input)
if args.mixup:
ce_loss = mixup_criterion(criterion, output, target_a, target_b,
lam, args.smooth)
else:
ce_loss = criterion(output, target, smooth=args.smooth)
mixed_loss = ce_loss
if args.admm:
admm.z_u_update(args, ADMM, model, device, train_loader, optimizer,
epoch, input, i, writer) # update Z and U
ce_loss, admm_loss, mixed_loss = admm.append_admm_loss(
args, ADMM, model, ce_loss) # append admm loss
if args.admm_mask:
admm.y_k_update(args, ADMM, model, device, train_loader, optimizer,
epoch, input, i, writer) # update Y\K
ce_loss, admm_loss, mixed_loss = admm.append_mask_loss(
args, ADMM, model, mixed_loss)
# measure accuracy and record loss
acc1, _ = accuracy(output, target, topk=(1, 5))
losses.update(ce_loss.item(), input.size(0))
top1.update(acc1[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
if args.admm or args.admm_mask:
mixed_loss.backward(retain_graph=True)
else:
ce_loss.backward()
if pre_mask:
with torch.no_grad():
for name, W in (model.named_parameters()):
# shared layers
if name in args.fixed_layer:
W.grad *= 0
continue
# pruned weight layers: fix weight for previous task
if name in args.pruned_layer and name in pre_mask:
W.grad *= pre_mask[name].cuda()
# adaptively learn the mask: fix mask for trainable weight part
if args.adaptive_mask and 'mask' in name and args.admm:
W.grad *= args.mask[name.replace('w_mask',
'weight')].cuda()
#W.grad *= 100
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.log_interval == 0:
for param_group in optimizer.param_groups:
current_lr = param_group['lr']
print('({0}) lr:[{1:.5f}] '
'Epoch: [{2}][{3}/{4}]\t'
'Status: admm-[{5}] retrain-[{6}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f}% ({top1.avg:.3f}%)\t'.format(
args.optmzr,
current_lr,
epoch,
i,
len(train_loader),
args.admm,
args.masked_retrain,
batch_time=data_time,
loss=losses,
top1=top1))
if i % 100 == 0:
idx_loss_dict[i] = losses.avg
