Unable to reproduce results from the paper

[MLMonkATGY]

Hi. Can the exact code from run_eval.py be used to reproduce the results from Table 16 ? I tried to benchmark distill-whisper-v2 on distil-whisper/common_voice_13_0 dataset and found the WER is a few percent higher than what was reported in the paper?

[bryanyzhu]

+1

[sanchit-gandhi]

Hey @MLMonkATGY! Could you share the arguments you're passing to run_eval.py so that I can reproduce locally? I believe this is because we are using the BasicNormalizer in the PyTorch script run_eval.py: https://github.com/huggingface/distil-whisper/blob/3490d8e366b09ef5c89bf0162f283240a1a5d725/training/run_eval.py#L545-L548

Whereas in the original Flax scripts, we always used the EnglishNormalizer: https://github.com/huggingface/distil-whisper/blob/3490d8e366b09ef5c89bf0162f283240a1a5d725/training/flax/run_eval.py#L728

You should be able to reproduce the results one-to-one if you use the Flax script. I'll also update the PyTorch script to use the EnglishNormalizer if the language used is English!

[MLMonkATGY]

I used the following arguments for run_eval.py.

python run_eval.py \ --model_name_or_path "distil-whisper/distil-large-v2" \ --dataset_name distil-whisper/common_voice_13_0 \ --dataset_config_name en \ --dataset_split_name test \ --text_column_name text \ --batch_size 128 \ --dtype "bfloat16" \ --generation_max_length 256 \ --language "en" \ --attn_implementation "flash_attention_2" \ --streaming True

[sanchit-gandhi]

Hey @MLMonkATGY, after merging #132, I evaluated the model with the following:

#!/bin/bash

python run_eval.py \
    --model_name_or_path "distil-whisper/distil-large-v2" \
    --dataset_name "distil-whisper/common_voice_13_0" \
    --dataset_config_name "en" \
    --dataset_split_name "test" \
    --text_column_name "text" \
    --batch_size 128 \
    --dtype "bfloat16" \
    --generation_max_length 256 \
    --language "en" \
    --streaming True

And got a WER of 13.0%: https://wandb.ai/sanchit-gandhi/distil-whisper-speed-benchmark/runs/7qihyqbx?nw=nwusersanchitgandhi

This is within 0.1% of the 12.9% WER reported in the paper. This 0.1% difference is expected, since the paper WER results are in Flax on TPU, whereas the run_eval.py script is in PyTorch on GPU. There's an inherent difference in how matrix multiplications are implemented in both, giving a subtle difference in results. Note that all WER results from the paper are in Flax, so the comparison between large-v2 and distil-large-v2 is valid. Note also that all RTF values in the paper were computed in PyTorch on GPU, such that they're most applicable to downstream use cases. I hope that helps!

[sanchit-gandhi]

All in all, the PR #132 should now mean that evaluating models in English with the PyTorch script run_eval.py gives WER results that are within 0.1% of the WER results quoted in the paper (using the Flax script flax/run_eval.py).

[MLMonkATGY]

Thanks !