pad missing the choice to fill at start or end

[ingted]

In PyTorch:

import torch.nn.functional as F
source = torch.arange(10).unfold(0,3,1)
result = F.pad(input=source, pad=(0,0,1,0), mode='constant', value=0)
source, result

We can decide to pad start, end or both.

However in TorchSharp:

let torchAfterUnfoldT = (afterUnfoldT.move Backend.Torch).toTorch() //a tensor
let pad = torch.nn.ConstantPad1d(1L, 0.0)
let padded = pad.forward(torchAfterUnfoldT)
let pdata = padded.data<System.Single>().ToArray()

It always pads both start and end.

PS. with pad in DiffSharp, it pads start + end as well

afterUnfoldT.pad(seq [0; 0; 1])

[NiklasGustafsson]

Hmm, don't know what to say. In both the Module.ConstantPad1d and the functional.pad() case, we just pass on the data to the native library.

Have you tried it with functional.pad() in TorchSharp, i.e. not using the ConstantPad1d version?

[ingted]

I found it is great that functional.pad works... However the behavior is a little different from python version.

[ingted]

Hi @NiklasGustafsson,

Excuse me, do you think this difference is a correct one? If so, I would like to close this issue...^^a

[NiklasGustafsson]

Well, since the output formatting is different, I find it hard to tell. Given that it's all in the backend, and we don't do anything but pass the data from .NET to C++, I don't see why it wouldn't be.

However, @dayo05 is adding the capability to format tensor.ToString() using the Numpy style, which should make it easier to compare. Maybe we can keep this open until then.

[kaiidams]

@NiklasGustafsson There are two functions in PyTorch torch.nn.functional.pad() and torchvision.transforms.functional.pad(). TorchSharp.torch.nn.functional.pad() works like torchvision.transoforms.functional.pad(). Probably we want to move TorchSharp.torch.nn.functional.pad() to TorchSharp.torchvision.transforms.functional.pad() and add new TorchSharp.torch.nn.functional.pad().

import torchvision.transforms.functional as F
source = torch.arange(10).unfold(0,3,1)
result = F.pad(img=source, padding=(1,2), padding_mode='constant', fill=0)
source, result

(tensor([[0, 1, 2],
         [1, 2, 3],
         [2, 3, 4],
         [3, 4, 5],
         [4, 5, 6],
         [5, 6, 7],
         [6, 7, 8],
         [7, 8, 9]]),
 tensor([[0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0],
         [0, 0, 1, 2, 0],
         [0, 1, 2, 3, 0],
         [0, 2, 3, 4, 0],
         [0, 3, 4, 5, 0],
         [0, 4, 5, 6, 0],
         [0, 5, 6, 7, 0],
         [0, 6, 7, 8, 0],
         [0, 7, 8, 9, 0],
         [0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0]]))

import torch.nn.functional as F
source = torch.arange(10).unfold(0,3,1)
result = F.pad(input=source, pad=(1,2), mode='constant', value=0)
source, result

(tensor([[0, 1, 2],
         [1, 2, 3],
         [2, 3, 4],
         [3, 4, 5],
         [4, 5, 6],
         [5, 6, 7],
         [6, 7, 8],
         [7, 8, 9]]),
 tensor([[0, 0, 1, 2, 0, 0],
         [0, 1, 2, 3, 0, 0],
         [0, 2, 3, 4, 0, 0],
         [0, 3, 4, 5, 0, 0],
         [0, 4, 5, 6, 0, 0],
         [0, 5, 6, 7, 0, 0],
         [0, 6, 7, 8, 0, 0],
         [0, 7, 8, 9, 0, 0]]))

var source = torch.arange(10).unfold(0,3,1);
var result = torch.nn.functional.pad(input: source, pad: new long[] {1,2}, mode: PaddingModes.Constant, value: 0);
display(result)

[12x5], type = Int64, device = cpu
 0 0 0 0 0
 0 0 0 0 0
 0 0 1 2 0
 0 1 2 3 0
 0 2 3 4 0
 0 3 4 5 0
 0 4 5 6 0
 0 5 6 7 0
 0 6 7 8 0
 0 7 8 9 0
 0 0 0 0 0
 0 0 0 0 0

[NiklasGustafsson]

@kaiidams, thanks for that observation. That's exactly what we should do. I'll take care of it.

[NiklasGustafsson]

@kaiidams, @ingted -- if you want to take a look at PR #604 and review the changes that are meant to fix this issue, that would be great help.

[ingted]

Hi @NiklasGustafsson ,

I read the code change and it seems like it is a correct change.
However I don't have time to build the binary to test it.
If you want my feedback after I did an real execution, I will try to provide it on Sun.

PS. Sorry that I didn't investigate enough to know what @kaiidams noticed... Hope that didn't waste your time at first...

[NiklasGustafsson]

@ingted, there's no need to build and test. I was just looking for a code review. I have added unit tests that do show that the expected behavior is implemented. If the code looks good and doesn't miss any boundary cases, that's good enough.