Model Improvement

[PetrIvan]

Hi, the models were updated to be more accurate. The suggestions should now load faster due to adding WebGPU support and optimizing the list of suggested chords.

The Conditional Transformer S should now be almost as accurate as the Conditional Transformer L was, similar advancements were also made to the other models. See the table below for the current performance (check page 16 of the technical report for a comparison to the previous results).

name	accuracy	perplexity	parameters
Recurrent Network	60.21%	4.159	377,260
Transformer S	72.80%	2.496	529,803
Transformer M	75.87%	2.506	1,536,075
Transformer L	76.89%	2.417	3,426,059
Conditional Transformer S	74.98%	2.618	536,155
Conditional Transformer M	76.36%	2.468	1,553,859
Conditional Transformer L	76.66%	2.437	3,465,131

If you find any bugs or have any comments on the changes, please let me know!

Would you like to see more features or vote for the next ones? You can do so at Feature requests.