In this project, we aim to predict the next chord in Jazz music given the history of previous chords using LSTMs. Then, we study whether adding information of the melody boosts the performance compared to using only chord series. Our implementation uses PyTorch.
For more information on the project, please check our final report
This project is part of CS-433 Machine Learning @ EPFL.
Group Members:
- Nerea Carbonell
- Maximilian Gangloff
- Daniel Morales
├── data # Contains the file to preprocess the data
│ ├── combine_melody_beats.py # Combines melody with beats table
│ ├── dataset.py # Defines the torch Datasets
│ ├── multi_hot_encoding.py # Creates the different Datasets
│ ├── wjazzd.db # Original Weimar Jazz Database
│ ├── datasets # Folder to save the datasets when training/loading the models
├── models #
│ ├── trained_models #
│ | ├── 1_Baseline # Model containing only the encoded chords
│ | ├── 2_Melody # Encoded chords + Melody
│ | ├── 3_Bass # Encoded chords + Bass
│ | ├── 4_Melody_Bass # Encoded chords + Melody + Bass
│ | ├── 5_Melody_Duration # Encoded chords + Melody weighted with duration
│ | ├── 6_Melody_Weighted # Encoded chords + Melody weighted w/o duration
│ ├── lstm_chord_models.py # Models for the Baseline
│ ├── lstm_melody_models.py # Models for the Melody encoding
├── JazzChordPredictionLSTM_MND.pdf # Generated .pdf file of our report
└── README.md # README
└── cross_validation.py # Uses k-fold cross-validation to get average accuracy
└── helpers.py # Contains helper functions for the train and cross_validation file such as calculating the loss and accuracy
└── hp_search.py # Uses random or grid search to find best hyperparameters
└── load_model.py # Loads a selected model
└── qualitative_analysis.py # Analysis between Baseline and Melody
└── Tune Accuracy comparison.csv # Accuracy comparison of the Baseline and Melody for each test sample (song)
└── train.py # Trains a given model
To check the different models' accuracy, load the models and generate the datasets by running python run.py. There, every dataset and model will be loaded from its generated and saved .pickle and .pth files. Then, the accuracies will be calculated and displayed in the terminal resulting in the following table.
| Model | Test accuracy |
|---|---|
| Baseline | 50.89 % |
| Melody | 53.93 % |
| Melody weighted | 53.47 % |
| Melody duration | 53.26 % |
| Bass | 52.72 % |
| Melody + Bass | 53.68 % |
To train one of the models, from the command line run python train.py --dataset dataset_id, where dataset_id is the index preceding the different models listed in trained models. Also, if the dataset has been previously generated, add as an argument --use-saved-dataset True.
The following arguments can be passed to train.py:
| Argument | Description |
|---|---|
| --dataset | Integer between 1 and 6 to choose the model |
| --hidden-dim | Hidden layers, default is 192 which is the optimal for all models |
| --lstm-layers | LSTM layers, default is 2 which is the best for all models |
| --max-epochs | Maximum number of epochs |
| --early-stopping | Number of iteration before stopping after no improvement was found |
| --seed | Seed |
| --lr | Learning rate |
| --wd | Weight decay |
| --dropout | Dropout |
| --use-saved-dataset | Use saved dataset |
| --save-path | Path of the model to be saved |
To generate the plots and tables used in the qualitative analysis, run qualitative_analysis.py. The default models to compare are 1_Baseline and 2_Melody.