MIDISim Inference Endpoint
Custom endpoint handler for the MIDISim small pre-trained model, packaged for HuggingFace Inference Endpoints.
Accepts a base64-encoded MIDI file and returns a 512-dimensional float32 embedding vector via GPU inference (~2-3s on T4).
API
POST https://<endpoint-url>/
{
"inputs": "<base64-encoded MIDI bytes>"
}
Response:
[0.12, -0.43, 0.07, ..., 0.31]
(512 floats)
Environment variables
| Variable | Default | Description |
|---|---|---|
MIDISIM_MAX_SEQ_LEN |
1024 |
Token sequence length. Longer = more accurate but slower. |
Model
- Checkpoint:
midisim_small_pre_trained_model_2_epochs_43117_steps_0.3148_loss_0.9229_acc.pth - Architecture: Transformer encoder, depth=8, dim=512, heads=8
- Original repo: projectlosangeles/midisim
- Embeddings corpus: projectlosangeles/midisim-embeddings