changed to proper Xavier initialization, existing implementation was …

[eknag]

…resulting in a large negative bias, which was killing all gradients through the following relu. https://paperswithcode.com/method/xavier-initialization

[facebook-github-bot]

Hi @eknag!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!

[eknag]

Behavior before

 python run.py dlrm -d cuda -t train
Running train method from dlrm on cuda in eager mode with input batch size 2048 and precision fp32.
gen: tensor([[0.],
        [0.],
        [0.],
        ...,
        [0.],
        [0.],
        [0.]], device='cuda:0', grad_fn=<ReluBackward0>)
loss: 0.3434426486492157
derivative with respect to first linear layer: tensor([[0., 0., 0.,  ..., 0., 0., 0.],

[eknag]

Behavior after fix:

➜  benchmark git:(dev/eknag/fix_dlrm_bias_initialization) ✗ python run.py dlrm -d cuda -t train              
Running train method from dlrm on cuda in eager mode with input batch size 2048 and precision fp32.
gen: tensor([[0.0000],
        [0.0000],
        [0.0213],
        ...,
        [0.0000],
        [0.0231],
        [0.0000]], device='cuda:0', grad_fn=<ReluBackward0>)
loss: 0.3336436152458191
derivative with respect to first linear layer: tensor([[ 7.9563e-04,  7.3723e-04,  7.9948e-04,  ...,  7.8230e-04,

[xuzhao9]

@eknag Thanks for looking into this. The code is from upstream model: https://github.com/facebookresearch/dlrm/blob/main/dlrm_s_pytorch.py#L212

Could you please submit a PR for the facebookresearch/dlrm repository and see what is the feedback there? Thanks!

[eknag]

See facebookresearch/dlrm#358

[xuzhao9]

@aaronenyeshi Can you also help take a look?

@eknag Could you please also add the URL of the upstream PR (facebookresearch/dlrm#358) to the code? We would like to track all such upstream improvements in the code.

[eknag]

@xuzhao9 Added. Let me know if I should do anything else - I'm new to making contributions.

[xuzhao9]

LGTM

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[facebook-github-bot]

@xuzhao9 merged this pull request in 827f90b.