def get_dataset_stats(args):
args = copy.deepcopy(args)
modify_args(args)
transformers = transformer_factory.get_transformers(
args.image_transformers, args.tensor_transformers, args)
train, _, _ = dataset_factory.get_dataset(args, transformers, [])
data_loader = DataLoader(train,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=False,
collate_fn=ignore_None_collate)
means, stds = {0: [], 1: [], 2: []}, {0: [], 1: [], 2: []}
for batch in tqdm(data_loader):
tensor = batch['x']
if args.cuda:
tensor = tensor.cuda()
for channel in range(3):
tensor_chan = tensor[:, channel]
means[channel].append(torch.mean(tensor_chan))
stds[channel].append(torch.std(tensor_chan))
means = [torch.mean(torch.Tensor(means[channel])) for channel in range(3)]
stds = [torch.mean(torch.Tensor(stds[channel])) for channel in range(3)]
return means, stds
def get_image_to_right_side(args):
args = copy.deepcopy(args)
modify_args(args)
transformers = transformer_factory.get_transformers(args.image_transformers, args.tensor_transformers, args)
train, dev, test = dataset_factory.get_dataset(args, transformers, transformers)
image_to_side = {}
for dataset in [train,dev,test]:
data_loader = DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
drop_last=False,
pin_memory=True,
collate_fn=ignore_None_collate)
for batch in tqdm(data_loader):
img = batch['x']
paths = batch['path']
if args.cuda:
img = img.cuda()
B, C, H, W = img.size()
left_half = img[:, :, :, :W//2].contiguous().view(B,-1)
right_half = img[:, :, :, W//2:].contiguous().view(B,-1)
is_right_aligned = right_half.sum(dim=-1) > left_half.sum(dim=-1)
for indx, path in enumerate(paths):
image_to_side[path] = bool(is_right_aligned[indx])
return image_to_side
commit.hexsha, commit.message, commit.author, commit.committed_date))
if args.get_dataset_stats:
print("\nComputing image mean and std...")
args.img_mean, args.img_std = get_dataset_stats(args)
print('Mean: {}'.format(args.img_mean))
print('Std: {}'.format(args.img_std))
print("\nLoading data-augmentation scheme...")
transformers = transformer_factory.get_transformers(
args.image_transformers, args.tensor_transformers, args)
test_transformers = transformer_factory.get_transformers(
args.test_image_transformers, args.test_tensor_transformers, args)
# Load dataset and add dataset specific information to args
print("\nLoading data...")
train_data, dev_data, test_data = dataset_factory.get_dataset(args, transformers, test_transformers)
# Load model and add model specific information to args
if args.snapshot is None:
model = model_factory.get_model(args)
else:
model = model_factory.load_model(args.snapshot, args)
if args.replace_snapshot_pool:
non_trained_model = model_factory.get_model(args)
model._model.pool = non_trained_model._model.pool
model._model.args = non_trained_model._model.args
print(model)
# Load run parameters if resuming that run.
args.model_path = state.get_model_path(args)
print('Trained model will be saved to [%s]' % args.model_path)