Parameter-Efficient Fine-Tuning of Whisper-Large V2 in Colab on T4 GPU using 🤗 PEFT+INT8 training #988
Replies: 11 comments 27 replies
-
How do we get the timestamp using peft? |
Beta Was this translation helpful? Give feedback.
-
when finetuing, I get follow train error: |
Beta Was this translation helpful? Give feedback.
-
@pacman100 I flow https://github.com/huggingface/peft/blob/main/examples/int8_training/peft_bnb_whisper_large_v2_training.ipynb to do asr fine-tuning training on about 1w hours data, when run trainer.train(), I met follows errors: the recipe of finetuing maybe just for small train data, for large train data, it maybe has some problem, https://github.com/huggingface/peft/blob/main/examples/int8_training/peft_bnb_whisper_large_v2_training.ipynb my run codes:
|
Beta Was this translation helpful? Give feedback.
-
@pacman100 A different is that I use set_transfomer() to do feature_extractor on the fly, my dataset is very big, when I use map() to feature_extractor, my 2TB disk is full. |
Beta Was this translation helpful? Give feedback.
-
it seems not work, follow huggingface/peft#242 (comment)
|
Beta Was this translation helpful? Give feedback.
-
@pacman100 But can finetuning whisper retain its generalization ability or will it try to overfit to your training data? I have been observing that if I finetune huggingface multilingual whisper on lang1 and test WER in another language, its performance degrades in comparison to the original multilingual model. Aren't you only utilizing the language-specific tokens during finetuning and masking others? |
Beta Was this translation helpful? Give feedback.
-
How to load from a checkpoint? |
Beta Was this translation helpful? Give feedback.
-
Can not reference on windows. Because bitsandbytes not support windows. |
Beta Was this translation helpful? Give feedback.
-
For multi-GPU, finetune is very slow for 10,000 hours of data, only training 0.5 epochs a month |
Beta Was this translation helpful? Give feedback.
-
Attention ASR developers and researchers! 🚀 Great news, with the latest update of 🤗 PEFT, you can now fine-tune your Whisper-large model faster than ever before! The new update allows you to fit 5X larger batches with less than 10GB GPU VRAM, thanks to LoRA and @Tim_Dettmers's bnb packaged nicely in 🤗 PEFT. And the best part? You get a comparable WER, but just faster!! ⚡️
But that's not all, you no longer have to compromise on the training speed to maintain WER. In fact, in our experiments with the Marathi language, the WER was comparable with full fine-tuning runs of Whisper-large.
With 🤗 PEFT, you can now train a Whisper-large v2 model in less than 8GB GPU VRAM! 📉 . Without 🤗 PEFT, you would experience OOM on a Colab T4, but not anymore! You can easily save on storage and port tiny checkpoints, ~63 MB compared to 6.7 GB fully fine-tuned model. 🐜
And that's not all! For low latency, you can convert the PEFT model to ONNX and use ORT using 🤗 Optimum. Start experimenting today and fine-tune your Whisper using PEFT+INT8 in Colab on a language of your choice! Join our Discord community to get involved in the conversation and discuss your results and questions. 🔬
Check out the Colab notebook examples and start your ASR development journey with 🤗 PEFT today!
Links:
Beta Was this translation helpful? Give feedback.
All reactions