Musical-Instrument-Classification (ONNX)

This is an ONNX version of Bhaveen/Musical-Instrument-Classification. It was automatically converted and uploaded using this Hugging Face Space.

Usage with Transformers.js

See the pipeline documentation for audio-classification: https://huggingface.co/docs/transformers.js/api/pipelines#module_pipelines.AudioClassificationPipeline

---

Musical Instrument Classification Model

This model is a fine-tuned version of facebook/wav2vec2-base-960h for musical instrument classification. It can identify 9 different musical instruments from audio recordings with high accuracy.

Model Description

• Model type: Audio Classification
• Base model: facebook/wav2vec2-base-960h
• Language: Audio (no specific language)
• License: MIT
• Fine-tuned on: Custom musical instrument dataset (200 samples for each class)

Performance

The model achieves excellent performance on the evaluation set after 5 epochs of training:

• Final Accuracy: 93.33%
• Final ROC AUC (Macro): 98.59%
• Final Validation Loss: 1.064
• Evaluation Runtime: 14.18 seconds
• Evaluation Speed: 25.39 samples/second

Training Progress

Supported Instruments

The model can classify the following 9 musical instruments:

1. Acoustic Guitar
2. Bass Guitar
3. Drum Set
4. Electric Guitar
5. Flute
6. Hi-Hats
7. Keyboard
8. Trumpet
9. Violin

Usage

Quick Start with Pipeline

from transformers import pipeline
import torchaudio

# Load the classification pipeline
classifier = pipeline("audio-classification", model="Bhaveen/epoch_musical_instruments_identification_2")

# Load and preprocess audio
audio, rate = torchaudio.load("your_audio_file.wav")
transform = torchaudio.transforms.Resample(rate, 16000)
audio = transform(audio).numpy().reshape(-1)[:48000]

# Classify the audio
result = classifier(audio)
print(result)

Using Transformers Directly

from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
import torchaudio
import torch

# Load model and feature extractor
model_name = "Bhaveen/epoch_musical_instruments_identification_2"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModelForAudioClassification.from_pretrained(model_name)

# Load and preprocess audio
audio, rate = torchaudio.load("your_audio_file.wav")
transform = torchaudio.transforms.Resample(rate, 16000)
audio = transform(audio).numpy().reshape(-1)[:48000]

# Extract features and make prediction
inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(predictions, dim=-1)

print(f"Predicted instrument: {model.config.id2label[predicted_class.item()]}")

Training Details

Dataset and Preprocessing

• Custom dataset with audio recordings of 9 musical instruments
• Train/Test Split: 80/20 using file numbering (files < 160 for training)
• Data Balancing: Random oversampling applied to minority classes
• Audio Preprocessing:
  • Resampling to 16,000 Hz
  • Fixed length of 48,000 samples (3 seconds)
  • Truncation of longer audio files

Training Configuration

# Training hyperparameters
batch_size = 1
gradient_accumulation_steps = 4
learning_rate = 5e-6
num_train_epochs = 5
warmup_steps = 50
weight_decay = 0.02

Model Architecture

• Base Model: facebook/wav2vec2-base-960h
• Classification Head: Added for 9-class classification
• Parameters: ~95M trainable parameters
• Features: Wav2Vec2 audio representations with fine-tuned classification layer

Technical Specifications

• Audio Format: WAV files
• Sample Rate: 16,000 Hz
• Input Length: 3 seconds (48,000 samples)
• Model Framework: PyTorch + Transformers
• Inference Device: GPU recommended (CUDA)

Evaluation Metrics

The model uses the following evaluation metrics:

• Accuracy: Standard classification accuracy
• ROC AUC: Macro-averaged ROC AUC with one-vs-rest approach
• Multi-class Classification: Softmax probabilities for all 9 instrument classes

Limitations and Considerations

1. Audio Duration: Model expects exactly 3-second audio clips (truncates longer, may not work well with shorter)
2. Single Instrument Focus: Optimized for single instrument classification, mixed instruments may produce uncertain results
3. Audio Quality: Performance depends on audio quality and recording conditions
4. Sample Rate: Input must be resampled to 16kHz for optimal performance
5. Domain Specificity: Trained on specific instrument recordings, may not generalize to all variants or playing styles

Training Environment

• Platform: Google Colab
• GPU: CUDA-enabled device
• Libraries:
  • transformers==4.28.1
  • torchaudio==0.12
  • datasets
  • evaluate
  • imblearn

Model Files

The repository contains:

• Model weights and configuration
• Feature extractor configuration
• Training logs and metrics
• Label mappings (id2label, label2id)

---

Model trained as part of a hackathon project