Cuda support for affine transforms

[siarez]

Seems like affine transforms don't support Cuda. Or am I missing something? Because this code

import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
from torchsample import transforms, TensorDataset

my_transforms = transforms.RandomAffine(rotation_range=180,
                 translation_range=0.2,
                 shear_range=None,
                 zoom_range=(0.8, 1.2))

test_imgs = torch.from_numpy(my_images).float().cuda()  # removing cuda() gets rid of the error
test_dataset = TensorDataset(test_imgs, input_transform=my_transforms)
test_loader = DataLoader(test_dataset, batch_size=1, shuffle=False)

for _data in tqdm(test_loader, total=len(test_loader)):
    print(index)
    plot_sample(_data.squeeze_().cpu())
    index += 1

Gives this error:

  0%|          | 0/4 [00:00<?, ?it/s]Traceback (most recent call last):
  File "/home/sia/Desktop/radar-image-recognition/data_manipul.py", line 70, in <module>
    for _data in tqdm(test_loader, total=len(test_loader)):
  File "/home/sia/anaconda3/lib/python3.6/site-packages/tqdm/_tqdm.py", line 953, in __iter__
    for obj in iterable:
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 179, in __next__
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 179, in <listcomp>
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/datasets.py", line 250, in __getitem__
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/datasets.py", line 250, in <listcomp>
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/transforms/affine_transforms.py", line 92, in __call__
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/transforms/affine_transforms.py", line 130, in __call__
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/utils.py", line 131, in th_affine2d
  File "/home/sia/anaconda3/lib/python3.6/site-packages/torchsample-0.1.3-py3.6.egg/torchsample/utils.py", line 174, in th_bilinear_interp2d
TypeError: gather received an invalid combination of arguments - got (int, !torch.LongTensor!), but expected (int dim, torch.cuda.LongTensor index)

After a bit of debugging, I think I get this error because the transform tensors are in CPU while my data is in GPU. Is there a way to move transform tensors to GPU? Or do I have to hack the code?

[RohitKeshari]

@siarez, @ncullen93 I do have the same error

`File "build/bdist.linux-x86_64/egg/torchsample/transforms/affine_transforms.py", line 322, in call File "build/bdist.linux-x86_64/egg/torchsample/utils.py", line 131, in th_affine2d

File "build/bdist.linux-x86_64/egg/torchsample/utils.py", line 174, in th_bilinear_interp2d

TypeError: gather received an invalid combination of arguments - got (int, torch.LongTensor), but expected (int dim, torch.cuda.LongTensor index) `

[amirid]

@siarez @RohitKeshari I added the following lines to the torchsample/functions/affine.py file and it works.

if cuda: coords = Variable(th_iterproduct(x.size(1),x.size(2),x.size(3)).float(), requires_grad=False).cuda() else: coords = Variable(th_iterproduct(x.size(1),x.size(2),x.size(3)).float(), requires_grad=False) Find the function which is rising the error and add .cuda() to the variable which is expected to be cuda tensor. Set the variable cuda True if you want to run it on GPU.

[mbchang]

It would be helpful to integrate cuda support more deeply into torchsample. For example, in torchsample/transforms/affine_transforms.py, instead of

        zoom_matrix = th.FloatTensor([[zx, 0, 0],
                                      [0, zy, 0],
                                      [0, 0,  1]])

if we directly initialize it as

        zoom_matrix = th.cuda.FloatTensor([[zx, 0, 0],
                                      [0, zy, 0],
                                      [0, 0,  1]])

we can avoid the overhead of converting a cpu Tensor to gpu.