How to debug torch.nn.export()

1

Hi,

I’m trying to convert a Pytorch model to CoreML format. As suggested, I’m aproaching this issue in two steps:

  • Pytorch to ONNX
  • ONNX to CoreML format

However, while converting a simple UNET with ResNet backend to ONNX, I get the following error:
RuntimeError: Failed to export an ONNX attribute, since it’s not constant, please try to make things (e.g., kernel size) static if possible

I took a look at torch.onnx code, but it is not easy to find exactly what line of the code (model) is provoking this error.

So, my question is: how to debug this conversion? There is a way to find what is causing this error? (turning verbose to True do not help)

Regards,

2

You can put prints everywhere you want to debug. The ONNX serializer traces the model, so it will execute it once, take a look at the documentation for more details.

Here is a snippet showing how you could debug:

import torch
class Conv(torch.nn.Module):
	def __init__(self):
		super(Conv, self).__init__()
		self.conv = torch.nn.Conv2d(3, 7, 5, 1, 2)
	def forward(self, x):
		print("-" * 40)
		print("DEBUG forward BEFORE convolution")
		print("-" * 40)
		x = self.conv(x)
		print("-" * 40)
		print("DEBUG forward AFTER convolution")
		print("-" * 40)
		return x

model = Conv()
dummy_input = torch.rand(16, 3, 100, 100)
torch.onnx.export(model, dummy_input, "debug.onnx", verbose=True)
2 Likes
3

Thanks @LeviViana,

I did not know that it would work. Looking at the call stack the execution was not anywhere near my model (it was inside torch.onnx in many calls). I took a look at the onnx code and found out where I could print the name of the current operation (even if symbolic), so I did that using gdb as I usually do.

In the end I found out a bug that is already registered at github.

Regards