composer

composer copied to clipboard

This is pretty important since training the Transformer-decoder model is very memory-intensive. Currently, due to the lack of distributed training support, to train the model, we must reduce the hyperparameters...

galacticglum

Beam search generation will dramatically improve the results of the model.

galacticglum

Currently, both the MusicRNN and Transformer models share nearly identical training loops. It would nice if we could refactor the bulk of this code into a general function.

galacticglum

Add option to plot the attention weights of the Transformer model during training.

galacticglum

Currently, all values in the config file have to be defined. The CLI won't try to infer default values if they are not defined in the config. The project provides...

galacticglum