README.md

---
library_name: onnx
tags:
  - onnx
  - optical-music-recognition
  - musicxml
  - omr
---

# homr ONNX checkpoints

This repository is a **mirror** of the ONNX checkpoints that
[liebharc/homr](https://github.com/liebharc/homr) downloads at runtime, published here
so that client-side (browser) builds can fetch them.

It exists because the models on homr's GitHub Release tag are served without
`Access-Control-Allow-Origin`, which blocks a browser from downloading them. Hugging
Face serves these files with CORS enabled. See `HuggingFaceModels.md` in the homr
repository for the full rationale and workflow.

**The source of truth is the GitHub release tag**, not this repository:

```
https://github.com/liebharc/homr/releases/download/onnx_checkpoints/
```

If the two ever disagree, the GitHub release wins.

## Models

Three models, each in fp32 and fp16:

| File | Role |
| --- | --- |
| `segnet_308-3296ccd40960f90ca6ab9c035cca945675d30a0f.onnx` | Segmentation (staff lines, note heads, stems/rests, bar lines, clefs/keys) |
| `segnet_308-3296ccd40960f90ca6ab9c035cca945675d30a0f_fp16.onnx` | …fp16 |
| `encoder_pytorch_model_465-597144cab54c8f6d0f6c9619df5c5312694eadd6.onnx` | TrOMR transformer encoder (ConvNeXt) |
| `encoder_pytorch_model_465-597144cab54c8f6d0f6c9619df5c5312694eadd6_fp16.onnx` | …fp16 |
| `decoder_pytorch_model_465-597144cab54c8f6d0f6c9619df5c5312694eadd6.onnx` | TrOMR transformer decoder, with KV cache |
| `decoder_pytorch_model_465-597144cab54c8f6d0f6c9619df5c5312694eadd6_fp16.onnx` | …fp16 |

Exact sizes:

| File | Size |
| --- | --- |
| `segnet_308-….onnx` | 54.66 MB |
| `segnet_308-…_fp16.onnx` | 27.34 MB |
| `encoder_pytorch_model_465-….onnx` | 50.41 MB |
| `encoder_pytorch_model_465-…_fp16.onnx` | 25.24 MB |
| `decoder_pytorch_model_465-….onnx` | 45.12 MB |
| `decoder_pytorch_model_465-…_fp16.onnx` | 89.61 MB |

Note that **fp16 is not smaller for the decoder** (89.61 MB vs 45.12 MB). A browser
build targeting WebGPU should use fp16 segnet + fp16 encoder + fp32 decoder
(97.7 MB total) rather than all-fp16 (292.4 MB).

## Usage in a browser

```js
const url =
  'https://huggingface.co/ngbcoder/Homr-onnx/resolve/main/' +
  'segnet_308-3296ccd40960f90ca6ab9c035cca945675d30a0f_fp16.onnx';

const session = await ort.InferenceSession.create(url);
```

Cache the bytes (`caches.open(...)` / Cache Storage) keyed by the full URL: at ~98 MB
per visitor, re-downloading on every page load is not acceptable.

Prefer to pin a specific revision? Replace `main` with a commit SHA from the
**Files** tab of this repo.

## Intended input/output shapes

| Model | Input | Output |
| --- | --- | --- |
| segnet | `[batch, 3, 320, 320]` float32 or float16 (matches the fp16 variant) | `[batch, 6, 320, 320]` logits, `argmax` over axis 0 |
| encoder | `[1, 1, H, W]` float32/float16, H ≤ 256, W ≤ 1280 | `context` |
| decoder | `rhythms`, `pitchs`, `lifts`, `articulations`, `slurs`, `context`, `cache_len`, `cache_in0..31` | `out_rhythms`, `out_pitchs`, `out_lifts`, `out_positions`, `out_articulations`, `out_slurs`, `attention`, `cache_out0..31` |

The decoder expects `32` KV-cache tensors because it has `decoder_depth = 8` layers with
`4` cache tensors each.

Normalisation for the encoder input is `(pixel / 255 - 0.7931) / 0.1738`, with a single
channel.

## Provenance

These weights come from homr's training pipeline, which builds on:

- the segmentation models of [oemer](https://github.com/BreezeWhite/oemer)
- the transformer model of [Polyphonic-TrOMR](https://github.com/NetEase/Polyphonic-TrOMR)

homr itself is licensed **AGPL-3.0**. Please confirm the licensing terms of the
upstream model weights before redistributing them.

## Updating

Re-export from homr's training pipeline, publish to the `onnx_checkpoints` release tag,
then mirror here. Never change a model in only one of the two places.