# -------------------------------------------------------
# Set up the training routine
network = nn.DataParallel(
model.float(),
device_ids=device_ids).to(device)
val_dataloader = torch.utils.data.DataLoader(
dataset=OmniDepthDataset(
root_path=input_dir,
path_to_img_list=val_file_list),
batch_size=1,
shuffle=True,
num_workers=num_workers,
drop_last=False)
trainer = OmniDepthTrainer(
experiment_name,
network,
None,
val_dataloader,
None,
None,
checkpoint_dir,
device,
validation_sample_freq=validation_sample_freq)
trainer.evaluate_upright(checkpoint_path, num_tests, rot_range, device, seed)
0.068,
]
else:
assert True, 'Unsupported network type'
# Make the checkpoint directory
mkdirs(checkpoint_dir)
# -------------------------------------------------------
# Set up the training routine
network = nn.DataParallel(model.float(), device_ids=device_ids).to(device)
val_dataloader = torch.utils.data.DataLoader(dataset=OmniDepthDataset(
root_path=input_dir, path_to_img_list=val_file_list),
batch_size=1,
shuffle=False,
num_workers=num_workers,
drop_last=False)
trainer = OmniDepthTrainer(experiment_name,
network,
None,
val_dataloader,
None,
None,
checkpoint_dir,
device,
validation_sample_freq=validation_sample_freq)
trainer.evaluate(checkpoint_path)
num_workers=num_workers,
drop_last=False)
criterion = MultiScaleL2Loss(alpha_list, beta_list)
# Set up network parameters with Caffe-like LR multipliers
param_list = set_caffe_param_mult(network, lr, 0)
optimizer = torch.optim.Adam(params=param_list, lr=lr)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=step_size,
gamma=lr_decay)
trainer = OmniDepthTrainer(experiment_name,
network,
train_dataloader,
val_dataloader,
criterion,
optimizer,
checkpoint_dir,
device,
visdom=[vis, env],
scheduler=scheduler,
num_epochs=num_epochs,
validation_freq=validation_freq,
visualization_freq=visualization_freq,
validation_sample_freq=validation_sample_freq,
num_samples=num_samples)
trainer.train(checkpoint_path, load_weights_only)