Itai Allouche • Adam katav

• MusicGenreClassification
  • Background
  • The Model
  • Dataset
  • Agenda
  • Results
    • 30s model
    • 15s model
    • 10s model
  • Docker
  • Training
    • Train 30s model
  • Run the model
    • Run the model - from huggingface 🤗
    • Run the model - using python

For our final project in the Technion DL course (046211), we chose to classify music genres over the GTZAN dataset.
The approach to the problem is using a pre-trained Wav2Vec2 transformer model.
As appose to most existing models, a transformer will use the raw time series data which is the reason we predicted an improvement over existing methods

We used the model facebook/wav2vec2-large-100k-voxpopuli from huggingface, Facebooks Wav2Vec2 model pre-trained on 100k unlabeled subset of speech data.

We used the femiliar GTZAN dataset.
The dataset consists of 1000 audio tracks each 30 seconds long.
It contains 10 genres, each represented by 100 tracks:
The genres are: blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock
The tracks are all 22050Hz Mono 16-bit audio files in .wav format.

The model was trained on 30s tracks.
performance:
87% accuracy on validation set

77% accuracy on test set

75.5% accuracy on test set

The model was trained on 10s tracks. Each 30s track was divided into 3 sub-tracks 10s long
performance:

74.5% accuracy on test set

The project is intended to run in huggingface docker image
For instructions on how to install docker:
https://docs.docker.com/engine/install/

Replace train_30s_model.py with your chosen model

docker run --name gtzan --rm -it --ipc=host --gpus=all -v $PWD:/home huggingface/transformers-pytorch-gpu python3 /home/train_30s_model.py

This command spins up a docker container from the official huggingface image, mounts the repo directory and run the training script

Open the Model in hugging face.

Note that hugging face server supports tracks up to 2-3 minutes

docker run --name gtzan --rm -it --ipc=host --gpus=all -v $PWD:/home huggingface/transformers-pytorch-gpu

docker run --name gtzan --rm -it -v $PWD:/home huggingface/transformers-pytorch-gpu

In the container either use a python script file or via the interactive interpreter:

from transformers import pipeline
import torchaudio
import sys
MODEL_NAME = 'adamkatav/wav2vec2_100k_gtzan_30s_model'
SONG_IN_REPO_DIR_PATH = '/home/rolling_stones.wav'

pipe = pipeline(model=MODEL_NAME)
audio_array,sample_freq = torchaudio.load(SONG_IN_REPO_DIR_PATH)
resample = torchaudio.transforms.Resample(orig_freq=sample_freq)
audio_array = audio_array.mean(axis=0).squeeze().numpy()
output = pipe(audio_array)
print(output)