Models v1.1.0

All twelve directed part pairs, so every part can be both the input and the answer. Eight of these are new; the four from v1.0.0 are unchanged and carry the same weights and hashes.

input	Drums	Bass	Melody	Harmony
Drums	—	0.4208	0.5199	0.3051
Bass	0.5308	—	0.5397	0.3037
Melody	0.5405	0.4455	—	0.3027
Harmony	0.6017	0.4516	0.5675	—

Val loss after 8 epochs, all on lmdc_17243_2bar_4res at 32 steps and 3.3M parameters.

hits-encdec-models-1.1.0.zip holds one model.pt per run, which is all you need to run a model. hits-encdec-checkpoints-1.1.0.zip also carries the optimizer state for resuming training; its weights are verified identical to the ones in the other archive. Both include a MANIFEST.json with each run's parts, dataset, epochs, val loss and sha256.

The hosted interface at p3zo.github.io/rhythmic-relationships runs int8 ONNX exports of these same weights.

Models v1.0.0

The first four trained hits encoder-decoders, all on lmdc_17243_2bar_4res at 32 steps and 3.3M parameters, 8 epochs each.

pair	run	val loss
Melody → Bass	youthfulness_2608221606	0.4455
Bass → Melody	pian_2608230003	0.5397
Melody → Harmony	lota_2608230042	0.3027
Harmony → Bass	boloman_2608230049	0.4516

hits-encdec-models-1.0.0.zip holds one model.pt per run, which is all you need to run a model. hits-encdec-checkpoints-1.0.0.zip also carries the optimizer state, for resuming training; its weights are verified identical to the ones in the other archive. Both include a MANIFEST.json with each run's parts, dataset, epochs, val loss and sha256.

The hosted interface at p3zo.github.io/rhythmic-relationships runs int8 ONNX exports of these same weights, which live in docs/data/models and are not part of this archive.