module 'encoding._ext.encoding_lib' has no attribute 'Encoding_Float_sum_square_Forward'

[mario98]

Although compiled successfully (carefully following the instructions), or at least no error messages has been shown during compilation, it seems that the module is not installed correctly.

I get the following error message:

  File "/home/<user>/code/networks/extractors.py", line 259, in forward
    out = self.start_features(x)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/nn/modules/container.py", line 67, in forward
    input = module(input)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/nn/syncbn.py", line 43, in forward
    xsum, xsquare = self.sharedT(sum_square(input.unsqueeze(3)))
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/functions/syncbn.py", line 59, in sum_square
    return _sum_square.apply(input)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/functions/syncbn.py", line 28, in forward
    encoding_lib.Encoding_Float_sum_square_Forward(
AttributeError: module 'encoding._ext.encoding_lib' has no attribute 'Encoding_Float_sum_square_Forward'

extractor.py line 259 refers to extractor.py line 250 from a deep learning repo where I replaced nn.BatchNorm2d with encoding.nn.BatchNorm2d.

Indeed, the folder /home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/_ext/encoding_lib does not contain any .py files for me.

What could be the issue here?

I am using:
pytorch 0.3.1 (build from source), I cloned the pytorch repo at state 2b4748011b5881583567bb166801ca6625f2fdda which is associated with the 0.3.1 release. pytorch 0.4 is too unstable and has too many (partly unlisted) breaking changes. I had to revert to v0.3.1
conda 4.3 / cuda 8.0 / Edit: python 3.5

[zhanghang1989]

Are you running the script in the repo base folder? Try to run it in different folder.

[mario98]

No, this is not the issue here. I'm running my script from a separate folder, say /home/<user>/code. But thanks for having a look at this. I'll keep you updated when I found something that could be responsible for this issue.
Is it intentional that the folder /home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/_ext does not contain any .py files?

[zhanghang1989]

something wrong with the installation, there should be __init__.py automatically generated:
_encoding_lib.abi3.so* __init__.py __pycache__/

[mario98]

Ok, strange. It just didn't copy the __init__.py file for some reason. I also couldn't find anything in the log files, Everything seemed to be run though without errors. I tried installation multiple times.

I ended up just copying this file manually from the sources to this directory, this solved the issue.

But now I get:

  [...]
    my_loss.backward()
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/autograd/variable.py", line 167, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/autograd/__init__.py", line 99, in backward
    variables, grad_variables, retain_graph)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/autograd/function.py", line 91, in apply
    return self._forward_cls.backward(self, *args)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/encoding/functions/syncbn.py", line 47, in backward
    gradInput, input.data.view(B, C, -1), gradSum, gradSquare)
  File "/home/<user>/.conda/envs/pytorch-0.3.1-3.5-s/lib/python3.5/site-packages/torch/utils/ffi/__init__.py", line 180, in safe_call
    result = torch._C._safe_call(*args, **kwargs)
TypeError: initializer for ctype 'struct THCudaTensor *' must be a cdata pointer, not Variable

Is this related to my incomplete installation or actually a new bug?

P.S. cffi package version 1.11.5 is installed on my system.

Edit: Solved by changing two lines of code!
It was not related to this issue, it's rather a pytorch 0.3.1 compatibility issue, I opened a new issue report for this.

And finally, the code runs on my machine! Thanks so far for all the helpful comments!

[zhanghang1989]

Please checkout the newest version and see if you still have the trouble. See details in PR #51

[zhanghang1989]

fixed in PR #59