Mert Base

🚀 MERT

MERT (Acoustic Music Understanding Model with Large-Scale Self-supervised Training) is designed to address the challenges in audio classification by leveraging teacher models to provide pseudo labels during masked language modelling (MLM) style acoustic pre-training.

The pre-trained weights of MERT are sourced from m-a-p/MERT-v1-95M. In this repository, MERT is registered for the AutoModelForAudioClassification auto class.

✨ Features

• Incorporates teacher models for pseudo labels in MLM-style acoustic pre-training.
• Compatible with the AutoModelForAudioClassification auto class.

🚀 Quick Start

The following steps and code example demonstrate how to use MERT for audio classification.

💻 Usage Examples

Basic Usage

import numpy as np
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification

# Some configurations
model_id = 'yangwang825/mert-base'
batch_size = 4
num_classes = 10
max_duration = 1.0

# Initialise the extractor and model
feature_extractor = AutoFeatureExtractor.from_pretrained(
    model_id, 
    trust_remote_code=True
)
mert = AutoModelForAudioClassification.from_pretrained(
    model_id,
    num_labels=num_classes,
    ignore_mismatched_sizes=True,
    trust_remote_code=True
)

# Simulate a list of waveforms (e.g. four audio clips)
audio_arrays = [
    np.random.rand(16000, ),
    np.random.rand(24000, ),
    np.random.rand(22050, ),
    np.random.rand(44100, )
]
inputs = feature_extractor(
    audio_arrays, # List of waveforms in numpy array format
    sampling_rate=feature_extractor.sampling_rate, 
    max_length=int(feature_extractor.sampling_rate * max_duration), 
    padding='max_length', 
    truncation=True, 
    return_tensors='pt'
)
# The shape of `input_values` is (batch_size, sample_rate * max_duration)
input_values = inputs['input_values']
outputs = mert(**inputs)
# The shape of `logits` is (batch_size, num_classes)
logits = outputs['logits']