torch.backends.cudnn.deterministic = True
np.random.seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# fetch dataloaders
# fetch dataloaders
val_dl = dataloader.fetch_dataloader(args.data_dir, args.txt_val, "val",
params)
# Define the model
model = get_network(params).to(params.device)
#num_classes+1 for background.
metrics = OrderedDict({})
for metric in params.metrics:
metrics[metric] = get_metrics(metric, params)
# Reload weights from the saved file
model = utils.load_checkpoint(model,
is_best=True,
checkpoint_dir=args.checkpoint_dir)[0]
# Evaluate
eval_loss, val_metrics = evaluate(model,
val_dl,
loss_fn=None,
metrics=metrics,
params=params)
best_json_path = os.path.join(args.model_dir, "logs/evaluation.json")
for val_metric_name, val_metric_results in val_metrics.items():
ckpt_target_file_path = os.path.join(args.checkpoint_dir_target, best_ckpt_filename)
if os.path.exists(ckpt_target_file_path):
model_target = utils.load_checkpoint(model_target, ckpt_dir=args.checkpoint_dir_target, filename=best_ckpt_filename, is_best=True)[0]
print("=> loaded target model checkpoint form {}".format(ckpt_target_file_path))
else:
print("=> Initializing target model from scratch")
for p in model_source.parameters():
p.requires_grad = False
model_source.eval()
for p in model_target.parameters():
p.requires_grad = False
model_target.eval()
opt = optim.AdamW(transfer.parameters(), lr=params_transfer.learning_rate)
lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(
opt, max_lr=params_transfer.learning_rate, steps_per_epoch=len(train_dl), epochs=params_transfer.num_epochs, div_factor=20)
# fetch loss function and metrics
loss_fn = get_loss_fn(params_transfer)
# num_classes+1 for background.
metrics = OrderedDict({})
for metric in params_transfer.metrics:
metrics[metric] = get_metrics(metric, params_transfer)
# Train the model
logging.info("Starting training for {} epoch(s)".format(params_transfer.num_epochs))
train_and_evaluate(model_source, model_target, transfer, train_dl, val_dl_source, val_dl_target, opt, loss_fn, metrics,
params_transfer, lr_scheduler, args.checkpoint_dir_transfer, ckpt_filename, log_dir, writer)
None,
None,
ckpt_dir=args.checkpoint_dir_target,
filename=ckpt_filename)
print("=> loaded target model checkpoint form {}".format(
ckpt_target_file_path))
else:
print("=> Initializing target model from scratch")
# fetch loss function and metrics
loss_fn1 = get_loss_fn(params_source)
loss_fn2 = get_loss_fn(params_target)
metrics_depth = OrderedDict({})
for metric in params_source.metrics:
metrics_depth[metric] = get_metrics(metric, params_source)
metrics_segmentation = OrderedDict({})
for metric in params_target.metrics:
metrics_segmentation[metric] = get_metrics(metric, params_target)
# Train the model
logging.info("Starting training for {} epoch(s)".format(
params_transfer.num_epochs))
train_and_evaluate(
model_source, model_target, transfer, train_dl_all, val_dl_all,
val_dl_target, opt1, opt2, opt3, loss_fn1, loss_fn2, metrics_depth,
metrics_segmentation, params_transfer, lr_scheduler1, lr_scheduler2,
lr_scheduler3, args.checkpoint_dir_source, args.checkpoint_dir_target,
args.checkpoint_dir_transfer, ckpt_filename, log_dir, writer)