Intel/whisper-large-onnx-int4-inc

INT4 Whisper large ONNX Model

Whisper is a pre-trained model for automatic speech recognition (ASR) and speech translation. Trained on 680k hours of labelled data, Whisper models demonstrate a strong ability to generalise to many datasets and domains without the need for fine-tuning. This is the repository of INT4 weight only quantization for the Whisper large model in ONNX format, powered by Intel® Neural Compressor and Intel® Extension for Transformers.

This INT4 ONNX model is generated by Intel® Neural Compressor's weight-only quantization method.

Model Detail	Description
Model Authors - Company	Intel
Date	October 8, 2023
Version	1
Type	Speech Recognition
Paper or Other Resources	-
License	Apache 2.0
Questions or Comments	Community Tab

Intended Use	Description
Primary intended uses	You can use the raw model for automatic speech recognition inference
Primary intended users	Anyone doing automatic speech recognition inference
Out-of-scope uses	This model in most cases will need to be fine-tuned for your particular task. The model should not be used to intentionally create hostile or alienating environments for people.

Export to ONNX Model

The FP32 model is exported with openai/whisper-large:

optimum-cli export onnx --model openai/whisper-large whisper-large-with-past/ --task automatic-speech-recognition-with-past --opset 13

Install ONNX Runtime

Install onnxruntime>=1.16.0 to support MatMulFpQ4 operator.

Run Quantization

Build Intel® Neural Compressor from master branch and run INT4 weight-only quantization.

The weight-only quantization cofiguration is as below:

dtype	group_size	scheme	algorithm
INT4	32	sym	RTN

We provide the key code below. For the complete script, please refer to whisper example.

from neural_compressor import quantization, PostTrainingQuantConfig
from neural_compressor.utils.constant import FP32

model_list = ['encoder_model.onnx', 'decoder_model.onnx', 'decoder_with_past_model.onnx']
for model in model_list:
    config = PostTrainingQuantConfig(
        approach="weight_only",
        calibration_sampling_size=[8],
        op_type_dict={".*": {"weight": {"bits": 4, 
                                        "algorithm": ["RTN"], 
                                        "scheme": ["sym"], 
                                        "group_size": 32}}},)
    q_model = quantization.fit(
        os.path.join("/path/to/whisper-large-with-past", model), # FP32 model path
        config,
        calib_dataloader=dataloader)
    q_model.save(os.path.join("/path/to/whisper-large-onnx-int4", model)) # INT4 model path

Evaluation

Operator Statistics

Below shows the operator statistics in the INT4 ONNX model:

Model	Op Type	Total	INT4 weight	FP32 weight
encoder_model	MatMul	256	192	64
decoder_model	MatMul	449	321	128
decoder_with_past_model	MatMul	385	257	128

Evaluation of wer

Evaluate the model on librispeech_asr dataset with below code:

import os
from evaluate import load
from datasets import load_dataset
from transformers import WhisperForConditionalGeneration, WhisperProcessor, AutoConfig
model_name = 'openai/whisper-large'
model_path = 'whisper-large-onnx-int4'
processor = WhisperProcessor.from_pretrained(model_name)
model = WhisperForConditionalGeneration.from_pretrained(model_name)
config = AutoConfig.from_pretrained(model_name)
wer = load("wer")
librispeech_test_clean = load_dataset("librispeech_asr", "clean", split="test")

from optimum.onnxruntime import ORTModelForSpeechSeq2Seq
from transformers import PretrainedConfig
model_config = PretrainedConfig.from_pretrained(model_name)
predictions = []
references = []
sessions = ORTModelForSpeechSeq2Seq.load_model(
            os.path.join(model_path, 'encoder_model.onnx'),
            os.path.join(model_path, 'decoder_model.onnx'),
            os.path.join(model_path, 'decoder_with_past_model.onnx'))
model = ORTModelForSpeechSeq2Seq(sessions[0], sessions[1], model_config, model_path, sessions[2])
for idx, batch in enumerate(librispeech_test_clean):
    audio = batch["audio"]
    input_features = processor(audio["array"], sampling_rate=audio["sampling_rate"], return_tensors="pt").input_features
    reference = processor.tokenizer._normalize(batch['text'])
    references.append(reference)
    predicted_ids = model.generate(input_features)[0]
    transcription = processor.decode(predicted_ids)
    prediction = processor.tokenizer._normalize(transcription)
    predictions.append(prediction)
wer_result = wer.compute(references=references, predictions=predictions)
print(f"Result wer: {wer_result * 100}")

Metrics (Model Performance):

Model	Model Size (GB)	wer
FP32	8.8	3.04
INT4	1.9	3.05