Updated CONTRIBUTING guide.

[vfdev-5]

Description:

• Added conda-forge channel, otherwise old pytorch nightly is proposed.

[datumbox]

LGTM, thanks!

[fmassa]

cc @seemethere should we update the installation instructions in the PyTorch website to take this into account?

[seemethere]

The conda installation instructions for Python 3.9 include a disclaimer about installing with conda-forge.

Is this going to be a hard requirement for torchvision now?

[fmassa]

Is this going to be a hard requirement for torchvision now?

I'm not even sure anymore why we are adding conda-forge as a dependency -- is it because of libjpeg / libpng ? Because ffmpeg is currently not a hard-dependency for torchvision, so if it's the reason why we are adding the conda-forge dependency we can probably remove it.

cc @andfoy

[vfdev-5]

@fmassa now, I'm not that sure even if using or not conda-forge channel we can properly install pytorch nightly with cudatoolkit=11.0. Doing that as

conda install pytorch cudatoolkit=11.0 -c pytorch-nightly -c conda-forge
> 
The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    ca-certificates-2020.12.5  |       ha878542_0         137 KB  conda-forge
    certifi-2020.12.5          |   py38h578d9bd_1         143 KB  conda-forge
    conda-4.9.2                |   py38h578d9bd_0         3.0 MB  conda-forge
    libuv-1.40.0               |       hd18ef5c_0         920 KB  conda-forge
    python_abi-3.8             |           1_cp38           4 KB  conda-forge
    pytorch-1.8.0.dev20201112  |py3.8_cuda11.0.221_cudnn8.0.3_0       713.5 MB  pytorch-nightly
    typing_extensions-3.7.4.3  |             py_0          25 KB  conda-forge
    ------------------------------------------------------------
                                           Total:       717.8 MB

Proposed pytorch is still 1.8.0.dev20201112. Same with conda install pytorch cudatoolkit=11.0 -c pytorch-nightly.

But using cudatoolkit=10.2 with or without conda-forge, proposes more recent version

conda install pytorch cudatoolkit=10.2 -c pytorch-nightly
> The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    cudatoolkit-10.2.89        |       hfd86e86_1       365.1 MB
    pytorch-1.9.0.dev20210216  |py3.8_cuda10.2_cudnn7.6.5_0       673.6 MB  pytorch-nightly
    typing_extensions-3.7.4.3  |     pyha847dfd_0          25 KB
    ------------------------------------------------------------
                                           Total:        1.01 GB

If we would like to install cudatoolkit=11.2, seems like there is no way to install packages without specifying conda-forge channel...

[andfoy]

Actually, conda-forge installations may lead to incompatibilities between libjpeg/jpegturbo, since PIL depends on the latter, and jpegturbo is not available in Windows AFAIK. Also there could be potential errors with FFmpeg

[datumbox]

I'm on Python 3.8 and I'm getting also quite old versions:

conda install pytorch torchvision cudatoolkit=11.0 -c pytorch-nightly -c conda-forge
>

Collecting package metadata (current_repodata.json): done
Solving environment: done

[...]

The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    pytorch-1.8.0.dev20210208  |py3.8_cuda110_cudnn8.0.5_0       625.0 MB  pytorch-nightly
    torchvision-0.9.0.dev20210211|       py38_cu110        25.8 MB  pytorch-nightly
    ------------------------------------------------------------

[...]

Where is the latest nightly?

[vfdev-5]

@datumbox try to install with cudatoolkit=11.2, may help

[datumbox]

You are right:

conda install pytorch torchvision cudatoolkit=11.2 -c pytorch-nightly -c conda-forge
>

[...]

    package                    |            build
    ---------------------------|-----------------
    pytorch-1.9.0.dev20210217  |py3.8_cuda11.2_cudnn8.1.0_0       911.2 MB  pytorch-nightly
    torchvision-0.9.0.dev20210221|       py38_cu112        26.0 MB  pytorch-nightly
    ------------------------------------------------------------

[...]