[TODO] Investigate equivalence tests

[ydshieh]

(add a lot of assignees just to make you informed and kept updated in the future. Don't hesitate to remove yourself if you think it's irrelevant)

Currently the PT/TF/Flax equivalence tests use 1e-5 as the tolerance for the absolute differences of outputs.

We see that these tests failed with a non-negligible (although not carefully defined) frequency.

Create this page to track a list of models to investigate.

• FlaxWav2Vec2ModelTest (2.2888184e-05 > 1e-5)

  • https://app.circleci.com/pipelines/github/huggingface/transformers/37363/workflows/a4b06424-0ba8-4fbc-9054-6ff52fbf8145/jobs/411654

• TFGPT2EncoderDecoderModelTest (0.001009281724691391 > 1e-3)

  • https://app.circleci.com/pipelines/github/huggingface/transformers/37358/workflows/43c12161-33d8-4df5-ba3c-3e62a4507ee7/jobs/411579
    • This also happens to TFBERTEncoderDecoderModelTest
    • This is caused by some sequence in a batch which gets all 0s as attention mask (generated by ids_tensor) - may happens on both encoder and decoder (especially after combining with the causal mask).
    • For TFBERTEncoderDecoderModelTest, the difference is smaller than TFGPT2EncoderDecoderModelTest (by a magnitude of 5x~10x) -> this is due to the last hidden states in GPT2 is after layer norm (not the case for BERT).
    • If we look the cross attention diff between PT/TF, it is clear that we have the same issue (both in the magnitude of 1e-3)
    • The encoder attention diff between PT/TF is in the magnitude of 5e-8: not very sure why this doesn't get much larger.
      • This is because PT/TF (at least in BERT) has different encoder_extended_attention_mask: 1e-4 vs 1e-9.

• TFViTMAEModelTest (1.013279e-05 > 1e-5)

  • https://app.circleci.com/pipelines/github/huggingface/transformers/37319/workflows/5adfba7a-d12b-4e1e-9a7a-e33c7d5fd6ee/jobs/411002

[gante]

Another one to add to this list: tests/funnel/test_modeling_funnel.py::FunnelModelTest::test_pt_tf_model_equivalence. I've been getting a failure in this one every other day -- example: https://app.circleci.com/pipelines/github/huggingface/transformers/38007/workflows/2a98b7b1-5ad0-4b80-a702-1887c620193f/jobs/421265

[ydshieh]

Another one to add to this list: tests/funnel/test_modeling_funnel.py::FunnelModelTest::test_pt_tf_model_equivalence. I've been getting a failure in this one every other day -- example: https://app.circleci.com/pipelines/github/huggingface/transformers/38007/workflows/2a98b7b1-5ad0-4b80-a702-1887c620193f/jobs/421265

Thanks. @stas00 also reported this. I will take a look~

[ydshieh]

Another one to add to this list: tests/funnel/test_modeling_funnel.py::FunnelModelTest::test_pt_tf_model_equivalence. I've been getting a failure in this one every other day -- example: https://app.circleci.com/pipelines/github/huggingface/transformers/38007/workflows/2a98b7b1-5ad0-4b80-a702-1887c620193f/jobs/421265

(just for the record) Among 500 runs:

• 34 runs have FunnelForMaskedLM.output.logits at around 1e-5 ~ 2e-5: so ~ 6.8% chance of failure 😢
• 66 runs at around 9e-6
• 38 runs at around 8e-6

(so > 25% to get close to 1e-5)

[gante]

@ydshieh I believe you can add the WIP label to stop the bot :)

[ydshieh]

I am afraid I will completely forget this issue. But if this brother you guys, OK for me. Thanks for the tip, I didn't know about it