gabprato / batch-transforms

Batch equivalent of PyTorch Transforms.

Batch Transforms

Batch equivalent of PyTorch Transforms.

transform_batch = transforms.Compose([
    ToTensor(),
    Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])

for images in data_iterator:
    images = transform_batch(images)
    output = model(images)

Normalize

Applies the equivalent of torchvision.transforms.Normalize to a batch of images.

Note: This transform acts out of place by default, i.e., it does not mutate the input tensor.

__init__(mean, std, inplace=False, dtype=torch.float, device='cpu')

• mean (sequence) – Sequence of means for each channel.
• std (sequence) – Sequence of standard deviations for each channel.
• inplace (bool,optional) – Bool to make this operation in-place.
• dtype (torch.dtype,optional) – The data type of tensors to which the transform will be applied.
• device (torch.device,optional) – The device of tensors to which the transform will be applied.

__call__(tensor)

• tensor (Tensor) – Tensor of size (N, C, H, W) to be normalized.

 

RandomCrop

Applies the equivalent of torchvision.transforms.RandomCrop to a batch of images. Images are independently transformed.

__init__(size, padding=None, device='cpu')

• size (int) – Desired output size of the crop.
• padding (int, optional) – Optional padding on each border of the image. Default is None, i.e no padding.
• device (torch.device,optional) – The device of tensors to which the transform will be applied.

__call__(tensor)

• tensor (Tensor) – Tensor of size (N, C, H, W) to be randomly cropped.

 

RandomHorizontalFlip

Applies the equivalent of torchvision.transforms.RandomHorizontalFlip to a batch of images. Images are independently transformed.

Note: This transform acts out of place by default, i.e., it does not mutate the input tensor.

__init__(p=0.5, inplace=False)

• p (float) – probability of an image being flipped.
• inplace (bool,optional) – Bool to make this operation in-place.

__call__(tensor)

• tensor (Tensor) – Tensor of size (N, C, H, W) to be randomly flipped.

 

ToTensor

Applies the equivalent of torchvision.transforms.ToTensor to a batch of images.

__init__()

__call__(tensor)

• tensor (Tensor) – Tensor of size (N, C, H, W) to be tensorized.

batch-transforms's Issues

Swapped width and height

Hi Gabriele, huge respect for what you managed to achieve here.
Your knowledge and expertise about tensor indexing truly impress as previously I couldn't imagine how random crop operations could be done in batches.
And after studying your code for a while I'm gradually coming to understand how such complex indexing works.

But one small thing here on line 123 of batchnorms.py really buffles me:

w, h = padded.size(2), padded.size(3)

I can't understand why the height and width variables are swapped. If this is indeed your intention, could you kindly shed some light on why it should be this way?

Not really an issue just saying thanks

Hi,

I used batch transforms from you in my code base for multi-style stylization, thanks for the implementation.

Alex

license?

Hi, what is the license of this repo (could you add a LICENSE file?)
(thanks for this repo! :) )