openai

/

whisper-large-v2

like 1.56k

You can try using the Flax fine-tuning script provided here: https://github.com/huggingface/transformers/tree/main/examples/flax/speech-recognition#whisper-model

You can try using the Flax fine-tuning script provided here: https://github.com/huggingface/transformers/tree/main/examples/flax/speech-recognition#whisper-model

Thanks for you reply. So can it be converted to pytorch again? Because I will use faster whisper at last.

Yes, you can convert any Transformers Flax model into PyTorch by passing the from_flax=True argument to from_pretrained:

from transformers import WhisperForConditionalGeneration

# load flax weights into pytorch
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v2", from_flax=True)

# save pytorch weights
model.save_pretrained("./output_dir")

See the docs for more details: https://huggingface.co/docs/transformers/main_classes/model#transformers.PreTrainedModel.from_pretrained.from_flax

I have tried the example script as you reference, but it can not execute successfully. My kaggle notebook is below:
https://www.kaggle.com/code/lukejacob/whisper-finetune-tpu

I have test it on GPU(T4x2), it is ok. But when I change to TPU v3x8, it crashes.

You can try using the Flax fine-tuning script provided here: https://github.com/huggingface/transformers/tree/main/examples/flax/speech-recognition#whisper-model

Hello, @sanchit-gandhi Thanks for you reply. I have tried the example script as you reference, but it can not execute successfully. My kaggle notebook is below:
https://www.kaggle.com/code/lukejacob/whisper-finetune-tpu

I have test it on GPU(T4x2), it is ok. But when I change to TPU v3x8, it crashes.

Hey @LukeJacob2023 - I think your notebook is private so I can't view it! What is the crash message? If you're able to make the notebook public I can take a look. If you're hitting OOMs, you might need to use the adafactor optimiser instead of adamw (1 optimiser state instead of 2): https://huggingface.co/sanchit-gandhi/large-v2-ls-ft/blob/9d49b40c54fe6e3bf458f05d2e955470ea8dc8d0/run_finetuning.py#L854

Hey @LukeJacob2023 - I think your notebook is private so I can't view it! What is the crash message? If you're able to make the notebook public I can take a look. If you're hitting OOMs, you might need to use the adafactor optimiser instead of adamw (1 optimiser state instead of 2): https://huggingface.co/sanchit-gandhi/large-v2-ls-ft/blob/9d49b40c54fe6e3bf458f05d2e955470ea8dc8d0/run_finetuning.py#L854

Sorry, it is ok now:
https://www.kaggle.com/code/lukejacob/whisper-finetune-tpu. it is not oom, I don't think tpu can oom.

Hello @sanchit-gandhi i am student in ML and i want to fine-tune whisper only specified for ICELANDIC , just for icelandic i mean i need a very good W.E.R for it , can you help me i need it for my last project to become an engineer