def init_optimizer(self):
def group_param_func(named_params):
base = {
'params': [(n, p) for n, p in named_params
if n.startswith("base_model")],
"lr":
config["lr"]
}
head = {
'params': [(n, p) for n, p in named_params
if not n.startswith("base_model")],
"lr":
config["lr_head"]
}
return [head, base]
self.optimizer = get_optimizer(self.model,
self.config,
group_param_func=group_param_func)
def init_optimizer(self):
self.optimizer = get_optimizer(self.model, self.config)
# init model
encoder_sizes = [28 * 28, 1000, 500, 250, 30]
decoder_sizes = [30, 250, 500, 1000, 28 * 28]
net = deep_autoencoder(encoder_sizes=encoder_sizes,
decoder_sizes=decoder_sizes,
activation=act).to(args.device)
# init dataloader
trainloader, testloader = get_dataloader(dataset=args.dataset,
train_batch_size=args.batch_size,
test_batch_size=256)
# init optimizer
optim_name = args.optimizer.lower()
tag = optim_name
optimizer = get_optimizer(optim_name, net, args)
# init lr scheduler
lr_scheduler = get_lr_scheduler(optimizer, args)
# init criterion
criterion = torch.nn.BCEWithLogitsLoss()
# init summary writter
log_dir = get_log_dir(optim_name, args)
if not os.path.isdir(log_dir):
os.makedirs(log_dir)
writer = SummaryWriter(log_dir)
# create output directory for visualized results
#visualization_dir = f"visuals/mnist/{optim_name}"
