"PyTorch convert function for op '_convolution_mode' not implemented"

[lauriebyrum]

🌱 Describe your Feature Request

Pytorch has added 'same' padding for convolutions, but if you try to convert such models to coreml, you get the error "PyTorch convert function for op '_convolution_mode' not implemented".

There is a workaround discussed in recent comments on pytorch/pytorch#3867, but it would be great if we didn't have to change our model code for this, especially since tf->coreml supports this padding.

[lauriebyrum]

a little more of the story: i had a model that i had trained in tensorflow over a year ago. At that time, i used tfcoreml to convert it and had a lovely fast running model. Recently, i was trying to convert a new variant of the model to coreml and found it to be 2x as slow. This is because coremltools isn't dealing with conversion to hwc. SO: i had already mostly moved my model to pytorch, so I have been trying to get the year-old architecture to convert from pytorch to coreml and run as fast as the tfcoreml produced model. I have been having to work around this '_convolution_mode' issue in order to convert at all. Today, I used a giant sledgehammer and forced pytorch->coreml to respect same padding instead of working around it. Finally, my model runs fast. Without the sledgehammer, it runs 100ms slower!

Long story short: there are workarounds to this issue, but they come with a very expensive cost. Please implement this!

[lauriebyrum]

We tried adding "_convolution_mode" to the alias list here https://github.com/apple/coremltools/blob/main/coremltools/converters/mil/frontend/torch/ops.py#L561. Then it complained about having strings as input. I think it was expecting padding ints. "ValueError: Op "input.305" (op_type: conv) Input pad="input.305_pad_0" expects integer tensor but got tensor[2,str]"

[TobyRoseman]

I can reproduce this problem with the following code:

import coremltools as ct
import torch
from torch import nn

model = nn.Conv2d(1,1,3, padding="same")
x = torch.randn(1,1,3,3)
traced_model = torch.jit.trace(model.eval(), x)
ct_model = ct.convert(traced_model, inputs=[ct.TensorType(shape=x.shape)])

@lauriebyrum - thanks for reporting this issue. This does seem like a high priority issue. I'll look into getting at least some level of support for PyTorch convolutions with same padding.