Weird results.

[thrillerist]

Hi Olivia, I have run the notebooks, and get following output. The result here is not as good as your demo video. Any idea? Thanks.

[oawiles]

I assume you are just naively running the notebooks.

The problem here is probably a different version of pytorch. In the readme I have listed the version the models were trained in. It is important to use this exact version -- some defaults changed between versions of pytorch meaning that even pretrained models give different results.

As a way to debug, the output should match the images that you see when you first downloaded this github repo (or you can view online on the github repo).

Hope this helps.

[thrillerist]

thanks, it works.

[ghost]

@oawiles seems !pip install torch==0.2.0 is no longer available Could not find a version that satisfies the requirement torch==0.2.0 (from versions: 0.1.2, 0.1.2.post1, 0.3.1, 0.4.0, 0.4.1, 1.0.0, 1.0.1, 1.0.1.post2, 1.1.0) No matching distribution found for torch==0.2.0 is there a alternative?

[oawiles]

I think you're probably safest trying distributions near that (e.g. 0.3.1, 0.1.2). Then try running the notebook and see if you get the same results as the preloaded results. If you do then great! if not, then that version is not compatible.

[oawiles]

Something you could also try (which I haven't) is to change the upsampling (I think the defaults changed with align_corners at some point) which may have been why the pre-computed models don't work properly anymore.

[ghost]

@oawiles with torch==0.3.1 I get this error AttributeError: module 'torch' has no attribute 'device' and with 0.1.2 ERROR: Command "python setup.py egg_info" failed with error code 1 in /tmp/pip-install-x3n_q9sj/torch/

[oawiles]

I'm not sure of installation errors. For the no attribute device : is this in the code? -- you could just edit it.

[ghost]

@oawiles I get these results with torch==0.4.1

[oawiles]

Yes this is wrong and I think this is the reason why. So you can try to update the code for upsampling to the old way -- align_corners=True.

[oawiles]

(I have not tested this so am not sure but basically you should be able to replicate the defaults in the notebook)

[ghost]

@oawiles im still unclear of how I could go about testing this

[oawiles]

When you view the notebook on github you should see some default results. When you load in new modules (+ additional changes) you should get the same results.

[mrgloom]

For macos and python2, after manually fixing code to run models on cpu I was able to reproduce results in https://github.com/oawiles/X2Face/blob/master/UnwrapMosaic/Pose2Face.ipynb

Python2:
    virtualenv -p /usr/local/bin/python2 x2face_env_py2
    python2 --version
    Python 2.7.15
    pip install https://download.pytorch.org/whl/torch-0.2.0.post3-cp27-none-macosx_10_7_x86_64.whl
    python2 -c "import torch; print(torch.__version__)"
    0.2.0_3
    pip install torchvision==0.2.0 --no-deps --no-cache-dir
    python2 -c "import torchvision; print(torchvision.__version__)"
    0.2.0
    python2 -m pip install ipykernel
    python2 -m ipykernel install --user --name=x2face_env_py2

[tlatlbtle]

Same issue. I fixed it by install following packages with specific versions: python=2.7

pip2 install https://download.pytorch.org/whl/cu75/torch-0.2.0.post3-cp27-cp27mu-manylinux1_x86_64.whl
pip2 install torchvision==0.2.0