backwarping for multiple GPUs

[ahmadmughees]

Hi, this is an amazing work and saved me from alot of work. I just want to ask that why you have written backwarping outside the Network class while you are also defining preprocess and basic inside the class. is there any theoretical background which I am missing right now?

Note. Actually I am trying to use your code to run on multiple GPUS and this function is creating a problem.

thanks.

return torch.nn.functional.grid_sample(input=tensorInput, grid=(Backward_tensorGrid[str(tensorFlow.size())] + tensorFlow).permute(0, 2, 3, 1),\
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1!

[sniklaus]

Thank you for your kind words! The location of the code doesn't matter much. I see it more as a utility function which is why I prefer it to be separate from the model. Regarding your error message, try changing the backward warping to the following.

def backwarp(tenInput, tenFlow):
	if str(tenFlow.shape) not in backwarp_tenGrid:
		tenHor = torch.linspace(-1.0 + (1.0 / tenFlow.shape[3]), 1.0 - (1.0 / tenFlow.shape[3]), tenFlow.shape[3]).view(1, 1, 1, -1).expand(-1, -1, tenFlow.shape[2], -1)
		tenVer = torch.linspace(-1.0 + (1.0 / tenFlow.shape[2]), 1.0 - (1.0 / tenFlow.shape[2]), tenFlow.shape[2]).view(1, 1, -1, 1).expand(-1, -1, -1, tenFlow.shape[3])

		backwarp_tenGrid[str(tenFlow.shape)] = torch.cat([ tenHor, tenVer ], 1)
	# end

	tenFlow = torch.cat([ tenFlow[:, 0:1, :, :] / ((tenInput.shape[3] - 1.0) / 2.0), tenFlow[:, 1:2, :, :] / ((tenInput.shape[2] - 1.0) / 2.0) ], 1)

	return torch.nn.functional.grid_sample(input=tenInput, grid=(backwarp_tenGrid[str(tenFlow.shape)].cuda() + tenFlow).permute(0, 2, 3, 1), mode='bilinear', padding_mode='border', align_corners=False)
# end

Closing for now since this should do the trick. I am happy to reopen this issue if it still persists though, just let me know.

[ahmadmughees]

Thanks for your responce.

actually, by using .cuda() was pusing the tensor on to default GPU, and other tensors were on different GPUs. the solution for that is pushing all the tensors on the same GPU by specifying GPU device. .cuda(tenFlow.device)

return torch.nn.functional.grid_sample(input=tenInput, grid=(backwarp_tenGrid[str(tenFlow.shape)].cuda(tenFlow.device) + tenFlow).permute(0, 2, 3, 1), mode='bilinear', padding_mode='border', align_corners=False)

[sniklaus]

Sorry that I didn't catch this (I didn't test it ...), and thank you for sharing your findings!