averaged_model = build_model()
averaged_model.to(device)
averaged_model.load_state_dict(checkpoint["state_dict"])
for name, param in averaged_model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
return model, optimizer, ema
model = build_model()
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
criterion = GaussianLoss()
ema = ExponentialMovingAverage(args.ema_decay)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
global_step, global_epoch = 0, 0
load_step = args.load_step
log = open(os.path.join(args.log, '{}.txt'.format(args.model_name)), 'w')
state = {k: v for k, v in args._get_kwargs()}
if load_step == 0:
list_train_loss, list_loss = [], []
log.write(json.dumps(state) + '\n')
test_loss = 100.0
else:
averaged_model.load_state_dict(checkpoint["state_dict"])
for name, param in averaged_model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
return model, optimizer, ema
model = build_model()
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
# optimizer = optim.Adamax(model.parameters(), lr=args.learning_rate)
criterion = GaussianLoss()
ema = ExponentialMovingAverage(
args.ema_decay
) #Maintains moving averages of variables by employing an exponential decay.
for name, param in model.named_parameters():
if param.requires_grad: #Every Tensor has a flag: requires_grad that allows for fine grained exclusion of subgraphs from gradient computation and can increase efficiency.
ema.register(name, param.data)
global_step, global_epoch = 0, 0
load_step = args.load_step
log = open(os.path.join(args.log, '{}.txt'.format(args.model_name)), 'w')
state = {k: v for k, v in args._get_kwargs()}
if load_step == 0:
list_train_loss, list_loss = [], []
log.write(json.dumps(state) + '\n')
test_loss = 100.0