Code to generate:
from transformers import WhisperForConditionalGeneration, AutoProcessor
new_config_values = dict(
d_model = 16,
decoder_attention_heads = 4,
decoder_layers = 1,
encoder_attention_heads = 4,
encoder_layers = 1,
num_hidden_layers = 1,
ignore_mismatched_sizes=True,
)
original_model = WhisperForConditionalGeneration.from_pretrained('openai/whisper-tiny', **new_config_values)
original_model.save_pretrained('converted')
original_processor = AutoProcessor.from_pretrained('openai/whisper-tiny')
original_processor.save_pretrained('converted')
Followed by:
$ mkdir -p ./converted/onnx
$ optimum-cli export onnx -m ./converted ./converted/onnx --task automatic-speech-recognition-with-past
$ find ./converted/onnx -type f ! -name "*.onnx" -delete
Usage (Transformers.js)
If you haven't already, you can install the Transformers.js JavaScript library from NPM using:
npm i @huggingface/transformers
Example: Transcribe audio from a URL.
import { pipeline } from '@huggingface/transformers';
const transcriber = await pipeline('automatic-speech-recognition', 'Xenova/tiny-random-WhisperForConditionalGeneration');
const url = 'https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav';
const output = await transcriber(url);
Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using 🤗 Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).