The dataset contains the results presented in the paper titled Convolutional Generative Adversarial Network, via Transfer Learning, for Traditional Scottish Music Generation presented at the EvoMUSART 2021 conference. The folder structure is the following:
- 100 Epochs from Scratch: contains the results of the 20 simulations performed with the model trained from scratch. Each simulation was run for 100 epochs.
- 100 Epochs Transfer Learning: contains the results of the 20 simulations performed with the model trained using transfer learning. Each simulation was run for 100 epochs.
Inside each of these folder, there are the folders corresponding to each of the 20 simulations. For example 100 Epochs from Scratch/haggis_balancedDataset_100Epochs_1 contains the results of the first simulation. The structure of each of these folders is the following:
- samples: contains the results of the training process in .png and .mid formats
- logs: contains .log and .txt files with the messages outputted from the training process
- eval: contains the results of the training process in .npy format
- checkpoints: contains the trained model saved at the last 5 epochs.