import torch
from torch import nn
from torchviz import make_dot
from models.retinaface import RetinaFace
from data import cfg_mnetv2, cfg_re50
net = RetinaFace(cfg=cfg_re50)
x = torch.randn(1, 3, 224, 224).requires_grad_(True)
y = net(x)
vis_graph = make_dot(y, params=dict(list(net.named_parameters()) + [('x', x)]))
vis_graph.view(filename="viz/resnet50.gv")
name = k[7:] # remove `module.`
else:
name = k
new_state_dict[name] = v
net.load_state_dict(new_state_dict)
cudnn.benchmark = True
# okay now we want to add new layers
for params in net.parameters(): # set all layers to false
params.requires_grad = False
net.ClassHead = net._make_class_head(fpn_num=3, inchannels=cfg['out_channel']) # re-initiliaze the layers
net.BboxHead = net._make_bbox_head(fpn_num=5, inchannels=cfg['out_channel'])
for name, param in net.named_parameters():
print(name,param.shape)
# for name, param in net.named_parameters(): # util to print layers that are now trainable
# if param.requires_grad:
# print (name)
Plist = []
for params in net.parameters(): # stores parameters that will be updated in Plist
if params.requires_grad:
Plist.append(params)
if num_gpu > 1 and gpu_train: # now transfer net to gpu if possible
net = torch.nn.DataParallel(net).cuda()
else:
net = net.cuda()
def train(opt, train_dict, device, tb_writer=None):
log_dir = Path(tb_writer.log_dir) if tb_writer else Path(
train_dict['logdir']) / 'logs'
wdir = str(log_dir / 'weights') + os.sep
os.makedirs(wdir, exist_ok=True)
last = wdir + 'last.pt'
best = wdir + 'best.pt'
results_file = 'results.txt'
with open(log_dir / 'hyp.yaml', 'w') as f:
yaml.dump(train_dict, 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'
rank = opt.global_rank
init_seeds(2 + rank)
train_path = train_dict['train']
test_path = train_dict['val']
train_dict['weights'] = last if not train_dict['pretrain'] or (
train_dict['pretrain'] and
not os.path.exists(train_dict['weights'])) else train_dict['weights']
model = RetinaFace(train_dict, phase='Train')
pretrained = False
if os.path.exists(train_dict['weights']):
pretrained = True
logger('Loading resume network from ====>{}'.format(
train_dict['weights']))
state_dict = torch.load(train_dict['weights'], map_location=device)
# create new OrderedDict that does not contain `module.`
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in state_dict['model'].items():
head = k[:7]
if head == 'module.':
name = k[7:] # remove `module.`
else:
name = k
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in model.named_parameters():
v.requires_grad = True
if '.bias' in k:
pg2.append(v) # biases
elif '.weight' in k and '.bn' not in k:
pg1.append(v) # apply weight decay
else:
pg0.append(v) # all else
if train_dict['adam']:
optimizer = optim.Adam(pg0,
lr=train_dict['lr0'],
betas=(train_dict['momentum'],
0.999)) # adjust beta1 to momentum
else:
optimizer = optim.SGD(pg0,
lr=train_dict['lr0'],
momentum=train_dict['momentum'],
nesterov=True)
optimizer.add_param_group({
'params': pg1,
'weight_decay': train_dict['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
epochs = train_dict['epoch']
lf = lambda x: ((
(1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
plot_lr_scheduler(optimizer, scheduler, epochs)
# Resume
start_epoch, best_fitness = 0, 0.0
if pretrained:
# Optimizer
if state_dict['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer'])
best_fitness = state_dict['best_fitness']
# Results
if state_dict.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(state_dict['training_results']) # write results.txt
# Epochs
start_epoch = state_dict['epoch'] + 1
if epochs < start_epoch:
logger.info(
'%s has been trained for %g epochs. Fine-tuning for %g additional epochs.'
% (weights, state_dict['epoch'], epochs))
epochs += state_dict['epoch'] # finetune additional epochs
del ckpt, state_dict
if train_dict['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
if cuda and rank != -1:
model = DDP(model,
device_ids=[opt.local_rank],
output_device=(opt.local_rank))
# Trainloader
batch_size = train_dict['batch_size']
image_size = train_dict['image_size']
# dataloader, dataset = create_dataloader(train_path,image_size, batch_size, opt, hyp=train_dict, augment=True,
# rect=opt.rect, rank=rank,
# world_size=opt.world_size, workers=train_dict['workers'])
rgb_mean = (104, 117, 123) # bgr order
dataset = WiderFaceDetection(train_path, preproc(image_size, rgb_mean))
sampler = torch.utils.data.distributed.DistributedSampler(dataset)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=8,
sampler=sampler,
pin_memory=True,
collate_fn=detection_collate)
criterion = MultiBoxLoss(num_classes, 0.35, True, 0, True, 7, 0.35, False)
priorbox = PriorBox(train_dict, image_size=(image_size, image_size))
with torch.no_grad():
priors = priorbox.forward()
priors = priors.cuda()
for epoch in range(start_epoch, epochs):
if rank != -1:
dataloader.sampler.set_epoch(epoch)
pbar = enumerate(dataloader)
if rank in [-1, 0]:
pbar = tqdm(pbar) # progress bar
optimizer.zero_grad()
for i, (
images, targets
) in pbar: # batch -------------------------------------------------------------
with amp.autocast(enabled=cuda):
images = images.cuda()
targets = [anno.cuda() for anno in targets]
out = model(images)
optimizer.zero_grad()
loss_l, loss_c, loss_landm = criterion(
out, priors, targets) * opt.world_size
loss = cfg['loc_weight'] * loss_l + loss_c + loss_landm
loss.backward()
optimizer.step()
load_t1 = time.time()
batch_time = load_t1 - load_t0
eta = int(batch_time * (max_iter - iteration))
if rank in [-1, 0]:
print(
'Epoch:{}/{} || Epochiter: {}/{} || Iter: {}/{} || Loc: {:.4f} Cla: {:.4f} Landm: {:.4f} || LR: {:.8f} || Batchtime: {:.4f} s || ETA: {}'
.format(epoch, max_epoch, (iteration % epoch_size) + 1,
epoch_size, iteration + 1, max_iter,
loss_l.item(), loss_c.item(),
loss_landm.item(), lr, batch_time,
str(datetime.timedelta(seconds=eta))))
torch.save(net.state_dict(),
wdir + os.sep + '{}_Final.pth'.format(i))