`Unfold` operator unsupported when exporting to ONNX

Question

When running

torch.onnx.export(
    model,
    (im, im),
    'wow.onnx',
    verbose=True,
    opset_version=opset,
    input_names=['t0', 't1'],
    output_names=['output']
)

on my model. I get:

Unsupported: ONNX export of operator Unfold , input size not accessible. Please feel free to request support or submit a pull request on PyTorch GitHub.

Any idea which operation Unfold (https://pytorch.org/docs/stable/generated/torch.nn.Unfold.html) could be replaced by so that the export succeeds?

Answer 1

I faced the same issue and struggled with it for a long time. You have to go into the model that you are trying to convert to ONNX. And replace the portion that is using "Unfold" operator with a different implementation.

For example, in my case, here is the original code that I had:

    # do patching
    if self.padding_patch == 'end':
        z = self.padding_patch_layer(z)
    z = z.unfold(dimension=-1, size=self.patch_len, step=self.stride)                   # z: [bs x nvars x patch_num x patch_len]
    z = z.permute(0,1,3,2) 

Here is what I replaced it with:

    # do patching manually
    if self.padding_patch == 'end':
        z = self.padding_patch_layer(z)
    
    # Manually unfold the input tensor
    batch_size, n_vars, seq_len = z.size()
    patches = []
    for i in range(0, seq_len - self.patch_len + 1, self.stride):
        patches.append(z[:, :, i:i+self.patch_len])
    
    z = torch.stack(patches, dim=-1)  # z: [bs x nvars x patch_len x patch_num]

Doing that made the ONNX conversion successful. Note that the second implementation may be a little inefficient and slower because of the introduction of the for loop, but it definitely works.