print(colored('==> ', 'blue') + 'GLOBAL-Net created.')
# Optimizer
optimizer = \
optim.Adam(filter(lambda p: p.requires_grad, model.parameters()),
lr=args.lr,
weight_decay=args.weight_decay)
# Scheduler
weights_loss_coeffs = [0.32, 0.08, 0.02]
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[3, 30, 40, 50, 60],
gamma=0.1)
if args.pretrained:
# reload from pre_trained_model
model, optimizer, scheduler, start_epoch, best_val = load_checkpoint(model, optimizer, scheduler,
filename=args.pretrained)
# now individually transfer the optimizer parts...
for state in optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(device)
cur_snapshot = os.path.basename(os.path.dirname(args.pretrained))
else:
if not os.path.isdir(args.snapshots):
os.mkdir(args.snapshots)
cur_snapshot = args.name_exp
if not osp.isdir(osp.join(args.snapshots, cur_snapshot)):
os.makedirs(osp.join(args.snapshots, cur_snapshot))
with open(osp.join(args.snapshots, cur_snapshot, 'args.pkl'), 'wb') as f:
lr=cfg['training']['learning_rate'],
weight_decay=0)
'''
torch with weight decay is equivelent to tensorflow tfa.optimizers.AdamW
we used tf.train.AdamOptimizer, which does not have weight decay. however, i dont know what is eauivelent to
the regularisation function in layers of tensorflow
'''
if cfg['training']['load_pre_training']:
path_to_pretrained_model = cfg['training']['pre_trained_weights']
if not os.path.isfile(cfg['training']['pre_trained_weights']):
raise ValueError('The path to the model pre-trained weights you indicated does not exist !')
# reload from pre_trained_model
print('path to pretrained model is {}'.format(path_to_pretrained_model))
model, optimizer, start_epoch, best_val = load_checkpoint(model, optimizer, filename=path_to_pretrained_model)
# now individually transfer the optimizer parts...
for state in optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(device)
save_path = os.path.join(cfg['write_dir'], cfg['name_exp'])
else:
save_path = os.path.join(cfg['write_dir'], cfg['name_exp'])
# creates the directory to save the images and checkpoints
if not osp.isdir(save_path):
os.mkdir(save_path)
with open(osp.join(save_path, 'args.pkl'), 'wb') as f:
pickle.dump(args, f)