NotImplementedError when using _backend.SpatialFullConvolution in latest pytorch

ttoosi · August 15, 2019, 9:06pm

Hi,
For my code, I need to access _backend.SpatialFullConvolution_updateOutput, when I use it in an earlier version of PyTorch it works fine but in the newer version (‘1.3.0.dev20190814’) I get the following error:

_backend.SpatialFullConvolution_updateOutput(
File “/conda-envs/pytorch_tensorflow/lib/python3.6/site-packages/torch/_thnn/utils.py”, line 27, in __ getattr __
raise NotImplementedError

Is this behavior expected?

Thanks,
Tahereh

gchanan · September 27, 2019, 6:23pm

what do you use it for?

ttoosi · September 27, 2019, 7:17pm

Hi,

I am trying to access backward path parameters e.g. weights, like what is used for here:

github.com

willwx/sign-symmetry/blob/master/functional/af_conv2d_function.py

"""
A conv2d autograd Function that supports different feedforward and feedback weights
Uses the same C functions used in the legacy SpatialConvolution module to implement actual computation
References:
    - https://github.com/L0SG/feedback-alignment-pytorch/blob/master/lib/fa_linear.py
    - https://pytorch.org/docs/master/notes/extending.html
    - torch/legacy/nn/SpatialConvolution.py
"""

import torch.autograd as autograd
from torch._thnn import type2backend


class AsymmetricFeedbackConv2dFunc(autograd.Function):

    @staticmethod
    def forward(context, input, weight, weight_feedback, bias, stride, padding):
        _backend = type2backend[input.type()]
        input = input.contiguous()
        output = input.new()

This file has been truncated. show original

I was wondering if there is a substitute for this backend in the new version?

Thanks,
Tahereh

gchanan · September 30, 2019, 12:58am

I’m not 100% sure what your code is doing, but it looks like you could probably use a backward hook to achieve the same effect. “SpatialConvolutionMM” is the same as “thnn_conv2d” I believe.

ttoosi · September 30, 2019, 6:50pm

Thanks for your reply.
backward hook only gives me the gradients computed in each layer. I want to be able to write my backward function which does not assume backpropagation (symmetric weight). That’s why access to _backend.SpatialConvolutionMM_updateGradInput
is important for my code. I would like to have access to such a function in the new Pytorch if it is possible.

gchanan · September 30, 2019, 9:20pm

we don’t expose a function that does that anymore to Python. It’s still available in C++ (though not guaranteed to continue to exist), should be thnn_backward.

I’d probably just write your own version for this case, though, if this is the only updateGradInput variant you need.

HaFred · November 27, 2020, 10:05am

Hi, sorry to bring this up. But may I know how do you access either the torch._thnn.type2backend or SpatialConvolutionMM in pytorch >1.5 nowadays? I am facing the same issue here…
Thank you very much.

ttoosi · November 27, 2020, 12:41pm

Hi! Unfortunately, I don’t have access to torch._thnn.type2backend in the newer pytorch. I had to find other ways.