Description
This model is a distilled version of the Whisper large v2 model using decoder pruning. It is trained to give the same distribution as the teacher(large-v2) model using Distillation loss (KL loss) + CE Loss. The original model contains 32 decoder layers, whereas the distilled model contains only 8 layers and achieves 4.2% WER on the librispeech dataset with finetuning for just one epoch. The decoding speed of the model is 2x faster than vanilla large-v2 and 40% smaller in size.
Train on your data
accelerate launch student-teacher-distillation-streaming.py --freeze_encoder --keep_punctuation
--keep_case --teacher_model_name_or_path openai/whisper-large-v2 --student_model_name_or_path large-v2-2
--student_cache_dir large-v2-2 --output_dir whisper-large-v2-2-en-cv --data_cache_dir commonvoice
--teacher_cache_dir cache --student_cache_dir large-v2-2-en-cv --text_column sentence
--train_dataset_name mozilla-foundation/common_voice_13_0 --train_dataset_config_name en --train_split_name train
--validation_dataset_name mozilla-foundation/common_voice_13_0 --validation_dataset_config_name en
--validation_split_name test --max_val_samples 2000
Inference
>>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset
>>> # load model and processor
>>> processor = WhisperProcessor.from_pretrained("rsonavane/distil-whisper-large-v2-8-ls")
>>> model = WhisperForConditionalGeneration.from_pretrained("rsonavane/distil-whisper-large-v2-8-ls")
>>> model.config.forced_decoder_ids = None
>>> # load dummy dataset and read audio files
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> sample = ds[0]["audio"]
>>> input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features
>>> # generate token ids
>>> predicted_ids = model.generate(input_features)
>>> # decode token ids to text
>>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
['<|startoftranscript|><|en|><|transcribe|><|notimestamps|> Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.<|endoftext|>']
>>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
[' Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.']
Limitations
This experiment aimed to explore the effectiveness of decoder pruning and distillation in enhancing performance after training. The model acquires a similar internal representation of the English language as its teacher model, but with improved inference speed and efficiency for downstream tasks. Additionally, it can be fine-tuned for multiple languages, maintaining the original model's performance while reducing inference latency. There are other frameworks such as JAX that can help improve the same.