WIP: Resolve "Upgrade to pytorch-1.0.1"

Closes #17 (closed)

mentioned in issue #20 (closed)

added 1 commit

• 1c87ea37 - [recipe] changed hardcoded pytorch and torchvision version, replaced with jinja variables

Compare with previous version

added 1 commit

• 36581e3a - [recipe] typo

Compare with previous version

added 1 commit

• a65f6ffd - [recipe] removed jinja variable for h5py

Compare with previous version

@heusch it's better to pin everything here to make sure your packages don't break in future. Something like:

- {{ pin_compatible('numpy') }}
- {{ pin_compatible('docopt') }}
- {{ pin_compatible('pytorch') }}
- {{ pin_compatible('torchvision') }}
...

changed this line in version 4 of the diff

You should not use jinja variables in the run section. Instead you should put it in the host section and use pin_compatible in the run section.

Ok will do

changed this line in version 4 of the diff

added 1 commit

• 7a162672 - [recipe] replaced jinja variable with pin_compatible

Compare with previous version

You don't need to pin python and setuptools but pin everything else.

added 1 commit

• 54560aae - [recipe] fix pin_compatible stuff

Compare with previous version

added 1 commit

• b77b60b0 - [recipe] changed the requirements to be compliant to what is stated in bob.devtools doc ..

Compare with previous version

added 1 commit

• d5c12df4 - [recipe] removed jinja variable for h5py

Compare with previous version

added 1 commit

• ace6dfb0 - [recipe] removed pin_compatible for docopt

Compare with previous version

After a bunch of trials and errors (mostly) on pinning with jinja variables and pin_compatible stuff in the conda recipe (inlcuding what is advised here: https://www.idiap.ch/software/bob/docs/bob/bob.devtools/master/templates.html#conda-recipe), I decided to step back for a while .. Feel free to keep on trying ;)

Maybe @andre.anjos has a solution to this ?

I manually pinned the version in the master branch (it needs to pass the CI for reasons related to BATL) and got the following error in the CI:

Traceback (most recent call last):
  File "/local/builds/bob/bob.learn.pytorch/miniconda/conda-bld/bob.learn.pytorch_1556527043921/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/bin/train_cnn.py", line 7, in <module>
    from bob.learn.pytorch.scripts.train_cnn import main
  File "/local/builds/bob/bob.learn.pytorch/miniconda/conda-bld/bob.learn.pytorch_1556527043921/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/lib/python3.6/site-packages/bob/learn/pytorch/scripts/train_cnn.py", line 42, in <module>
    import torch
  File "/local/builds/bob/bob.learn.pytorch/miniconda/conda-bld/bob.learn.pytorch_1556527043921/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/lib/python3.6/site-packages/torch/__init__.py", line 102, in <module>
    from torch._C import *
ImportError: /local/builds/bob/bob.learn.pytorch/miniconda/conda-bld/bob.learn.pytorch_1556527043921/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/lib/python3.6/site-packages/torch/lib/libtorch.so.1: undefined symbol: nvrtcGetProgramLogSize

Apparently, it is related to the usage of the library on the CPU and/or GPU ...

There is a related issue (with a solution) on PyTorch's GitHub https://github.com/pytorch/pytorch/issues/14973. Basically, linking to libnvrtc.so and libcuda.so. I don't think it's possible to do that on the CI (or is it ?). A possible workaround is presented here: https://stackoverflow.com/questions/55665606/how-to-fix-importerror-home-lib-libtorch-so-1-undefined-symbol-nvrt, but requires pytorch packages that are currently not in conda defaults channels ...

So, how do we proceed with that ?

Ok, apparently the tentative workaround (dicussed via email and using pytorch-cpu) does not work either. I hence suggest to revert bob-devel back to pytorch 0.4.1, when everything was working smoothly ... And we should keep it until a new (and CI compatible) version of pytorch will be released in conda defaults channels

@ageorge @amohammadi @andre.anjos @tlaibacher What do you think ?

@heusch downgrading bob-devel is a lot of work. We have updated our source code to new dependencies. For example we had to change our code to update to click 7, ffmpeg 4, and tensorflow 1.13. If reverted, some of our packages will break.

What you guys really need is a conda package for bob.learn.pytorch that works with both pytorch and pytorch-cpu. Here it is: bob.learn.pytorch-0.0.4b0-py36ha615f63_34.tar.bz2

mentioned in issue #17 (closed)

closed