def create_opt(parameters, config):
if config.opt == "SGD":
optimizer = optim.SGD(parameters, lr=config.lr, weight_decay=config.l2)
elif config.opt == "Adam":
optimizer = optim.Adam(parameters,
lr=config.lr,
weight_decay=config.l2)
elif config.opt == "Adadelta":
optimizer = optim.Adadelta(parameters,
lr=config.lr,
rho=config.rho,
eps=config.eps,
weight_decay=config.l2)
elif config.opt == "Adagrad":
optimizer = optim.Adagrad(parameters,
lr=config.lr,
weight_decay=config.l2)
elif config.opt == "AdamW":
print("Using AdamW")
optimizer = AdamW(parameters, lr=config.lr, weight_decay=config.l2)
return optimizer
for k in model_t_loc.state_dict():
if k in loaded_dict and sd[k].size() == loaded_dict[k].size():
sd[k] = loaded_dict[k]
loaded_dict = sd
model_t_loc.load_state_dict(loaded_dict)
# named_parameters()包含网络模块名称 key为模型模块名称 value为模型模块值，可以通过判断模块名称进行对应模块冻结
for key, value in model_t_loc.named_parameters():
value.requires_grad = False
del loaded_dict
del sd
del checkpoint
if args.mode != "onlyT":
params = model_s.parameters()
optimizer = AdamW(params, lr=args.lr, weight_decay=args.weight_decay)
model_s, optimizer = amp.initialize(model_s, optimizer, opt_level="O0")
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[5, 11, 17, 23, 29, 33, 47, 50, 60, 70, 90, 110, 130, 150, 170, 180, 190], gamma=0.5)
else:
params = model_t.parameters()
optimizer = AdamW(params, lr=args.lr, weight_decay=args.weight_decay)
model_t, optimizer = amp.initialize(model_t, optimizer, opt_level="O0")
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[5, 11, 17, 23, 29, 33, 47, 50, 60, 70, 90, 110, 130, 150, 170, 180, 190], gamma=0.5)
if args.transfer and args.mode !='onlyT':
snap_to_load = 'res50_loc_{}_KD_best'.format(seed)
print("=> loading checkpoint '{}'".format(snap_to_load))
checkpoint = torch.load(path.join(models_folder, snap_to_load), map_location='cpu')
loaded_dict = checkpoint['state_dict']
sd = model_s.state_dict()
for k in model_s.state_dict():
batch_size=batch_size,
num_workers=5,
shuffle=True,
pin_memory=False,
drop_last=True)
val_data_loader = DataLoader(val_train,
batch_size=val_batch_size,
num_workers=5,
shuffle=False,
pin_memory=False)
model = SeResNext50_Unet_Loc().cuda()
params = model.parameters()
optimizer = AdamW(params, lr=0.00015, weight_decay=1e-6)
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[
15, 29, 43, 53, 65, 80, 90, 100,
110, 130, 150, 170, 180, 190
],
gamma=0.5)
seg_loss = ComboLoss({'dice': 1.0, 'focal': 10.0}, per_image=False).cuda()
best_score = 0
_cnt = -1
torch.cuda.empty_cache()