Finding the best approach to classify music genre in GTZAN database

Author: Tianhao Liu (20205784)

Email: [email protected] | [email protected]

In this project, I tried two approachs to classify the genres of songs. One is to build deeplearning model from scratch, using CNN, MLP, LSTM, and GRU. Beside this, I also tried to fine tune exsiting pre-trained models (vgg19, resNext, and SqueezeNet)

Train From Scratch

Quick Start

1. Download the GTZAN dataset to this project folder
2. Unzip the downloaded database, the name of the downloaded database should be Data
3. If the database is in somewhere else or with other names, please write the path to config file hparams.yaml. You need to modify the audio_dir and image_dir in it.
4. main.ipynb is all you need.
5. After training, a trained model will be saved in folder checkpoints, and its loss record will be saved in folder logs.

Model	Status
MLP	✅
CNN	✅
LSTM	✅
GRU	✅

Fine Tune

1. install all the packages declared in requirements.txt
2. I have fine tuned 3 pre-trained models, that are: resnext, vgg19, and squeezenet. You can find them in finetune-resnext.ipynb, finetune-vgg.ipynb, and finetune-squeezenet.ipynb respectively.

Model	Status
VGG	✅
ResNext	✅
SqueezeNet	✅

Techniques in Training

Feature	Status
Early Stop	✅
Batch training	✅
Checkpoint	✅
Log (loss)	✅
Train-test-split	✅
Evaluation	✅