README.md

---
library_name: onnx
tags: [onnx, electron-microscopy, segmentation, cellmap, webgpu, onnxruntime-web]
---

# CellMap ONNX web models

Small 3D U-Nets for FIB-SEM organelle segmentation (mitochondria, nuclei, ER), exported to ONNX for **in-browser** inference with
[onnxruntime-web](https://onnxruntime.ai/) in the [cellmap-flow serverless client](https://github.com/mzouink/cellmap-flow/tree/serverless).

Every file takes `input` float32 `[1,1,z,y,x]` and returns `output` float32 **raw logits** `[1,C,z',y',x']`
(apply sigmoid, or threshold at 0). Input normalization: `uint8 / 255 * 2 - 1`
(`[{"name":"MinMaxNormalizer","min_value":0,"max_value":255},{"name":"LambdaNormalizer","expression":"2*x-1"}]`).
The input block size is fixed per file (`_in<N>` = N³ voxels); valid convolutions, so the output is N−88 per axis.

Two graph layouts of the **same models** (same weights, same outputs up to float rounding):

- `setup_XX/…` — **recommended**: every Conv3d rewritten as Conv2d over a depth-folded batch (~16× faster on WebGPU, see below).
- `conv3d/setup_XX/…` — native Conv3d nodes, for runtimes with a fast 3D conv (e.g. onnxruntime CUDA/CPU, TensorRT).

## Models

| folder | target | channels | voxel (nm) | architecture | params | checkpoint | val F1 |
|---|---|---|---|---|---|---|---|
| `setup_54` | mito_affinities | mito_aff_z, mito_aff_y, mito_aff_x | 16 | ResUNet3DValid | 10.7M | 260000 | 0.958 |
| `setup_53` | mito_affinities | mito_aff_z, mito_aff_y, mito_aff_x | 16 | Unet3dIso | 5.6M | 340000 | 0.955 |
| `setup_56` | nuc | nuc | 128 | ResUNet3DValid | 10.7M | 350000 | 0.765 |
| `setup_55` | nuc | nuc | 128 | Unet3dIso | 5.6M | 330000 | 0.645 |
| `setup_61` | nuc | nuc | 256 | ResUNet3DValid | 10.7M | 20000 | 0.738 |
| `setup_67` | er | er | 8 | ResUNet3DValid | 10.7M | 20000 | 0.806 |

## Download links and speed (recommended 2D-rewrite layout)

Times are steady-state ms per block on WebGPU (Chrome, Apple M-series GPU, real EM; see benchmark below).
Lower ms is better; `out Mvox/s` = output voxels per second (higher is better). fp16 is 1.3–2.2× faster than fp32,
halves download and GPU memory, and changes ≤0.05% of `logit > 0` decisions. fp32 is exact and also runs on the WASM (CPU) fallback.
The largest block of each model gives the best throughput (88 voxels of context are lost per axis per block).

| model | target | voxel | in → out | fp32 | fp32 ms | fp16 | fp16 ms | fp16 speedup | out Mvox/s fp32 / fp16 |
|---|---|---|---|---|---|---|---|---|---|
| `setup_54` | mito_affinities (3 ch) | 16 nm | 148³ → 60³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in148_fp32.onnx) | 1203 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in148_fp16.onnx) | 905 | 1.33× | 0.18 / 0.24 |
| `setup_54` | mito_affinities (3 ch) | 16 nm | 132³ → 44³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in132_fp32.onnx) | 796 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in132_fp16.onnx) | 573 | 1.39× | 0.11 / 0.15 |
| `setup_54` | mito_affinities (3 ch) | 16 nm | 116³ → 28³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in116_fp32.onnx) | 516 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_54/setup_54_mito_affinities_16nm_in116_fp16.onnx) | 323 | 1.60× | 0.04 / 0.07 |
| `setup_53` | mito_affinities (3 ch) | 16 nm | 172³ → 84³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in172_fp32.onnx) | 1137 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in172_fp16.onnx) | 741 | 1.53× | 0.52 / 0.80 |
| `setup_53` | mito_affinities (3 ch) | 16 nm | 140³ → 52³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in140_fp32.onnx) | 533 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in140_fp16.onnx) | 336 | 1.59× | 0.26 / 0.42 |
| `setup_53` | mito_affinities (3 ch) | 16 nm | 124³ → 36³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in124_fp32.onnx) | 348 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_53/setup_53_mito_affinities_16nm_in124_fp16.onnx) | 157 | 2.22× | 0.13 / 0.30 |
| `setup_56` | nuc (1 ch) | 128 nm | 148³ → 60³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in148_fp32.onnx) | 1208 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in148_fp16.onnx) | 921 | 1.31× | 0.18 / 0.23 |
| `setup_56` | nuc (1 ch) | 128 nm | 132³ → 44³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in132_fp32.onnx) | 811 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in132_fp16.onnx) | 574 | 1.41× | 0.11 / 0.15 |
| `setup_56` | nuc (1 ch) | 128 nm | 116³ → 28³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in116_fp32.onnx) | 527 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_56/setup_56_nuc_128nm_in116_fp16.onnx) | 324 | 1.63× | 0.04 / 0.07 |
| `setup_55` | nuc (1 ch) | 128 nm | 172³ → 84³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in172_fp32.onnx) | 1080 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in172_fp16.onnx) | 758 | 1.42× | 0.55 / 0.78 |
| `setup_55` | nuc (1 ch) | 128 nm | 140³ → 52³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in140_fp32.onnx) | 506 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in140_fp16.onnx) | 319 | 1.59× | 0.28 / 0.44 |
| `setup_55` | nuc (1 ch) | 128 nm | 124³ → 36³ | [fp32 (23 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in124_fp32.onnx) | 331 | [fp16 (11 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_55/setup_55_nuc_128nm_in124_fp16.onnx) | 156 | 2.12× | 0.14 / 0.30 |
| `setup_61` | nuc (1 ch) | 256 nm | 148³ → 60³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in148_fp32.onnx) | 1220 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in148_fp16.onnx) | 900 | 1.36× | 0.18 / 0.24 |
| `setup_61` | nuc (1 ch) | 256 nm | 132³ → 44³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in132_fp32.onnx) | 809 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in132_fp16.onnx) | 574 | 1.41× | 0.11 / 0.15 |
| `setup_61` | nuc (1 ch) | 256 nm | 116³ → 28³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in116_fp32.onnx) | 514 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_61/setup_61_nuc_256nm_in116_fp16.onnx) | 322 | 1.60× | 0.04 / 0.07 |
| `setup_67` | er (1 ch) | 8 nm | 148³ → 60³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in148_fp32.onnx) | 1233 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in148_fp16.onnx) | 890 | 1.39× | 0.18 / 0.24 |
| `setup_67` | er (1 ch) | 8 nm | 132³ → 44³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in132_fp32.onnx) | 809 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in132_fp16.onnx) | 569 | 1.42× | 0.11 / 0.15 |
| `setup_67` | er (1 ch) | 8 nm | 116³ → 28³ | [fp32 (43 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in116_fp32.onnx) | 532 | [fp16 (22 MB)](https://huggingface.co/cellmap/onnx-web-models/resolve/main/setup_67/setup_67_er_8nm_in116_fp16.onnx) | 323 | 1.65× | 0.04 / 0.07 |

Native-3D versions of every file are under `conv3d/` with the same names (e.g. `https://huggingface.co/cellmap/onnx-web-models/resolve/main/conv3d/setup_54/setup_54_mito_affinities_16nm_in148_fp32.onnx`). On WebGPU they are ~16× slower (setup_54 148³ fp32: 20.9 s vs 1.2 s per block) and fp16 does not help there (17–25 s), because onnxruntime-web runs Conv3d with a naive shader; use them for onnxruntime CUDA/CPU or TensorRT.

## Compression variants

| variant | weights / compute | file vs fp32 | notes |
|---|---|---|---|
| `fp32` | fp32 / fp32 | 1× | reference; best accuracy, works on WebGPU and WASM (CPU) |
| `fp16` | fp16 / fp16 | ½ | halves download and activation memory; WebGPU only (poor fp16 support on WASM) |
| `int8w` | int8 per-channel / fp32 | ¼ | smaller download only (weights dequantized at load) |
| `int8w_fp16` | int8 per-channel / fp16 | ¼ | smallest download + fp16 activations |

## Graph rewrites for WebGPU

Both are exact (same math, same weights; verified against the PyTorch model):

- **Conv3d → Conv2d over a depth-folded batch.** Each 3D conv is computed as a sum of 2D convs, one per depth tap, on the
  volume reshaped to `[depth, C, H, W]`. onnxruntime-web runs 3D convs with a naive WebGPU shader but 2D convs with tuned
  matmul kernels, so this is ~16× faster (setup_54, 148³ block, Apple M-series GPU: 20.9 s → 1.2 s).
- **3D ConvTranspose → 1×1×1 conv + depth-to-space** (Unet3dIso decoders); the WebGPU EP has no 3D ConvTranspose.

Peak memory is dominated by activations, not weights: the largest single tensor is ~40×N³ (ResUNet) or ~32×(N−2)³ (Unet3dIso)
values, e.g. 259 MB fp16 / 519 MB fp32 at N=148. Pick a smaller block if your GPU's WebGPU `maxBufferSize` is lower.

## Browser WebGPU benchmark

Measured in Chrome 154 on macOS, Apple GPU (WebGPU `metal-3`), onnxruntime-web 1.27.0-dev.20260506; best of 3 runs after one warm-up run.

Real EM (jrc_mus-liver, public OpenOrganelle data) at each model's voxel size. `vs fp32 flip %` = voxels whose `logit > 0`
decision differs from the same model's fp32 file on WebGPU.

| model | block | variant | ms / block | output voxels / s | max abs vs PyTorch (sampled) | flip % vs fp32 |
|---|---|---|---|---|---|---|
| `setup_54` | 148³ | `fp32` | 1203 | 179,551 | 0.001 | — |
| `setup_54` | 148³ | `fp16` | 905 | 238,674 | 0.082 | 0.0 |
| `setup_54` | 148³ | `int8w` | 1216 | 177,632 | 0.106 | 0.0 |
| `setup_54` | 148³ | `int8w_fp16` | 879 | 245,734 | 0.153 | 0.0 |
| `setup_54` | 132³ | `fp32` | 796 | 107,015 | 0.0 | — |
| `setup_54` | 132³ | `fp16` | 573 | 148,663 | 0.072 | 0.0 |
| `setup_54` | 132³ | `int8w` | 1096 | 77,723 | 0.096 | 0.0 |
| `setup_54` | 132³ | `int8w_fp16` | 558 | 152,659 | 0.09 | 0.0 |
| `setup_54` | 116³ | `fp32` | 516 | 42,543 | 0.001 | — |
| `setup_54` | 116³ | `fp16` | 323 | 67,963 | 0.048 | 0.0 |
| `setup_54` | 116³ | `int8w` | 531 | 41,341 | 0.103 | 0.0 |
| `setup_54` | 116³ | `int8w_fp16` | 328 | 66,927 | 0.101 | 0.0 |
| `setup_53` | 172³ | `fp32` | 1137 | 521,288 | 0.001 | — |
| `setup_53` | 172³ | `fp16` | 741 | 799,870 | 0.071 | 0.0 |
| `setup_53` | 172³ | `int8w` | 1013 | 585,098 | 0.208 | 0.0 |
| `setup_53` | 172³ | `int8w_fp16` | 737 | 804,212 | 0.166 | 0.0 |
| `setup_53` | 140³ | `fp32` | 533 | 263,805 | 0.0 | — |
| `setup_53` | 140³ | `fp16` | 336 | 418,476 | 0.076 | 0.0 |
| `setup_53` | 140³ | `int8w` | 494 | 284,632 | 0.268 | 0.0 |
| `setup_53` | 140³ | `int8w_fp16` | 328 | 428,683 | 0.225 | 0.0 |
| `setup_53` | 124³ | `fp32` | 348 | 134,069 | 0.0 | — |
| `setup_53` | 124³ | `fp16` | 157 | 297,172 | 0.073 | 0.0 |
| `setup_53` | 124³ | `int8w` | 317 | 147,180 | 0.194 | 0.0 |
| `setup_53` | 124³ | `int8w_fp16` | 160 | 291,600 | 0.201 | 0.0 |
| `setup_56` | 148³ | `fp32` | 1208 | 178,808 | 0.001 | — |
| `setup_56` | 148³ | `fp16` | 921 | 234,528 | 0.2 | 0.0 |
| `setup_56` | 148³ | `int8w` | 1219 | 177,194 | 0.379 | 0.0 |
| `setup_56` | 148³ | `int8w_fp16` | 880 | 245,455 | 0.34 | 0.0 |
| `setup_56` | 132³ | `fp32` | 811 | 105,036 | 0.001 | — |
| `setup_56` | 132³ | `fp16` | 574 | 148,404 | 0.16 | 0.0 |
| `setup_56` | 132³ | `int8w` | 822 | 103,630 | 0.447 | 0.0 |
| `setup_56` | 132³ | `int8w_fp16` | 559 | 152,386 | 0.266 | 0.0 |
| `setup_56` | 116³ | `fp32` | 527 | 41,655 | 0.001 | — |
| `setup_56` | 116³ | `fp16` | 324 | 67,753 | 0.119 | 0.0 |
| `setup_56` | 116³ | `int8w` | 526 | 41,734 | 0.342 | 0.0 |
| `setup_56` | 116³ | `int8w_fp16` | 326 | 67,337 | 0.332 | 0.0 |
| `setup_55` | 172³ | `fp32` | 1080 | 548,800 | 0.001 | — |
| `setup_55` | 172³ | `fp16` | 758 | 781,931 | 0.117 | 0.0 |
| `setup_55` | 172³ | `int8w` | 989 | 599,296 | 0.558 | 0.0 |
| `setup_55` | 172³ | `int8w_fp16` | 736 | 805,304 | 0.65 | 0.0 |
| `setup_55` | 140³ | `fp32` | 506 | 277,881 | 0.001 | — |
| `setup_55` | 140³ | `fp16` | 319 | 440,777 | 0.131 | 0.0 |
| `setup_55` | 140³ | `int8w` | 481 | 292,324 | 0.531 | 0.0 |
| `setup_55` | 140³ | `int8w_fp16` | 336 | 418,476 | 0.512 | 0.0 |
| `setup_55` | 124³ | `fp32` | 331 | 140,955 | 0.001 | — |
| `setup_55` | 124³ | `fp16` | 156 | 299,077 | 0.103 | 0.0 |
| `setup_55` | 124³ | `int8w` | 315 | 148,114 | 0.314 | 0.0 |
| `setup_55` | 124³ | `int8w_fp16` | 160 | 291,600 | 0.402 | 0.0 |
| `setup_61` | 148³ | `fp32` | 1220 | 177,049 | 0.001 | — |
| `setup_61` | 148³ | `fp16` | 900 | 240,000 | 0.124 | 0.0019 |
| `setup_61` | 148³ | `int8w` | 1224 | 176,471 | 0.328 | 0.0032 |
| `setup_61` | 148³ | `int8w_fp16` | 869 | 248,562 | 0.362 | 0.0037 |
| `setup_61` | 132³ | `fp32` | 809 | 105,295 | 0.001 | — |
| `setup_61` | 132³ | `fp16` | 574 | 148,404 | 0.164 | 0.0 |
| `setup_61` | 132³ | `int8w` | 808 | 105,426 | 0.407 | 0.0 |
| `setup_61` | 132³ | `int8w_fp16` | 555 | 153,485 | 0.482 | 0.0 |
| `setup_61` | 116³ | `fp32` | 514 | 42,708 | 0.001 | — |
| `setup_61` | 116³ | `fp16` | 322 | 68,174 | 0.145 | 0.0 |
| `setup_61` | 116³ | `int8w` | 525 | 41,813 | 0.303 | 0.0 |
| `setup_61` | 116³ | `int8w_fp16` | 327 | 67,131 | 0.434 | 0.0 |
| `setup_67` | 148³ | `fp32` | 1233 | 175,182 | 0.0 | — |
| `setup_67` | 148³ | `fp16` | 890 | 242,697 | 0.03 | 0.0148 |
| `setup_67` | 148³ | `int8w` | 1221 | 176,904 | 0.042 | 0.0282 |
| `setup_67` | 148³ | `int8w_fp16` | 870 | 248,276 | 0.048 | 0.0356 |
| `setup_67` | 132³ | `fp32` | 809 | 105,295 | 0.0 | — |
| `setup_67` | 132³ | `fp16` | 569 | 149,708 | 0.035 | 0.0129 |
| `setup_67` | 132³ | `int8w` | 796 | 107,015 | 0.037 | 0.0399 |
| `setup_67` | 132³ | `int8w_fp16` | 556 | 153,209 | 0.047 | 0.0434 |
| `setup_67` | 116³ | `fp32` | 532 | 41,263 | 0.0 | — |
| `setup_67` | 116³ | `fp16` | 323 | 67,963 | 0.039 | 0.0501 |
| `setup_67` | 116³ | `int8w` | 529 | 41,497 | 0.043 | 0.0866 |
| `setup_67` | 116³ | `int8w_fp16` | 327 | 67,131 | 0.066 | 0.0729 |

## Fidelity vs. fp32 PyTorch (onnxruntime CPU)

One real EM crop per model (jrc_c-elegans-comma-1 at the model's voxel size). `flip %` = voxels whose `logit > 0` decision
differs from PyTorch. A consistency check, not a re-run of F1 validation.

| file | MB | in → out | max abs logit diff | mean abs diff | flip % |
|---|---|---|---|---|---|
| `setup_54/setup_54_mito_affinities_16nm_in148_fp32.onnx` | 42.9 | 148³ → 60³ | 0.0005 | 0.0001 | 0.0000 |
| `setup_54/setup_54_mito_affinities_16nm_in148_fp16.onnx` | 21.6 | 148³ → 60³ | 0.0142 | 0.0024 | 0.0009 |
| `setup_54/setup_54_mito_affinities_16nm_in148_int8w.onnx` | 10.9 | 148³ → 60³ | 0.1366 | 0.0198 | 0.0094 |
| `setup_54/setup_54_mito_affinities_16nm_in148_int8w_fp16.onnx` | 10.9 | 148³ → 60³ | 0.1139 | 0.0191 | 0.0080 |
| `setup_54/setup_54_mito_affinities_16nm_in132_fp32.onnx` | 42.9 | 132³ → 44³ | 0.0003 | 0.0000 | 0.0000 |
| `setup_54/setup_54_mito_affinities_16nm_in132_fp16.onnx` | 21.6 | 132³ → 44³ | 0.0149 | 0.0025 | 0.0000 |
| `setup_54/setup_54_mito_affinities_16nm_in132_int8w.onnx` | 10.9 | 132³ → 44³ | 0.1452 | 0.0203 | 0.0192 |
| `setup_54/setup_54_mito_affinities_16nm_in132_int8w_fp16.onnx` | 10.9 | 132³ → 44³ | 0.1202 | 0.0191 | 0.0160 |
| `setup_54/setup_54_mito_affinities_16nm_in116_fp32.onnx` | 42.9 | 116³ → 28³ | 0.0001 | 0.0000 | 0.0000 |
| `setup_54/setup_54_mito_affinities_16nm_in116_fp16.onnx` | 21.6 | 116³ → 28³ | 0.0134 | 0.0026 | 0.0015 |
| `setup_54/setup_54_mito_affinities_16nm_in116_int8w.onnx` | 10.9 | 116³ → 28³ | 0.1302 | 0.0213 | 0.0137 |
| `setup_54/setup_54_mito_affinities_16nm_in116_int8w_fp16.onnx` | 10.9 | 116³ → 28³ | 0.1188 | 0.0209 | 0.0076 |
| `setup_53/setup_53_mito_affinities_16nm_in172_fp32.onnx` | 22.5 | 172³ → 84³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_53/setup_53_mito_affinities_16nm_in172_fp16.onnx` | 11.3 | 172³ → 84³ | 0.0618 | 0.0072 | 0.0044 |
| `setup_53/setup_53_mito_affinities_16nm_in172_int8w.onnx` | 5.8 | 172³ → 84³ | 0.4799 | 0.0488 | 0.0380 |
| `setup_53/setup_53_mito_affinities_16nm_in172_int8w_fp16.onnx` | 5.8 | 172³ → 84³ | 0.5517 | 0.0509 | 0.0351 |
| `setup_53/setup_53_mito_affinities_16nm_in140_fp32.onnx` | 22.5 | 140³ → 52³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_53/setup_53_mito_affinities_16nm_in140_fp16.onnx` | 11.3 | 140³ → 52³ | 0.0618 | 0.0073 | 0.0009 |
| `setup_53/setup_53_mito_affinities_16nm_in140_int8w.onnx` | 5.8 | 140³ → 52³ | 0.3368 | 0.0430 | 0.0085 |
| `setup_53/setup_53_mito_affinities_16nm_in140_int8w_fp16.onnx` | 5.8 | 140³ → 52³ | 0.3255 | 0.0451 | 0.0092 |
| `setup_53/setup_53_mito_affinities_16nm_in124_fp32.onnx` | 22.5 | 124³ → 36³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_53/setup_53_mito_affinities_16nm_in124_fp16.onnx` | 11.3 | 124³ → 36³ | 0.0396 | 0.0069 | 0.0007 |
| `setup_53/setup_53_mito_affinities_16nm_in124_int8w.onnx` | 5.8 | 124³ → 36³ | 0.2672 | 0.0384 | 0.0129 |
| `setup_53/setup_53_mito_affinities_16nm_in124_int8w_fp16.onnx` | 5.8 | 124³ → 36³ | 0.2747 | 0.0395 | 0.0157 |
| `setup_56/setup_56_nuc_128nm_in148_fp32.onnx` | 42.9 | 148³ → 60³ | 0.0035 | 0.0003 | 0.0005 |
| `setup_56/setup_56_nuc_128nm_in148_fp16.onnx` | 21.6 | 148³ → 60³ | 0.0539 | 0.0065 | 0.0069 |
| `setup_56/setup_56_nuc_128nm_in148_int8w.onnx` | 10.9 | 148³ → 60³ | 0.5656 | 0.0669 | 0.0755 |
| `setup_56/setup_56_nuc_128nm_in148_int8w_fp16.onnx` | 10.9 | 148³ → 60³ | 0.4653 | 0.0600 | 0.0560 |
| `setup_56/setup_56_nuc_128nm_in132_fp32.onnx` | 42.9 | 132³ → 44³ | 0.0019 | 0.0002 | 0.0000 |
| `setup_56/setup_56_nuc_128nm_in132_fp16.onnx` | 21.6 | 132³ → 44³ | 0.0543 | 0.0064 | 0.0012 |
| `setup_56/setup_56_nuc_128nm_in132_int8w.onnx` | 10.9 | 132³ → 44³ | 0.4437 | 0.0661 | 0.0153 |
| `setup_56/setup_56_nuc_128nm_in132_int8w_fp16.onnx` | 10.9 | 132³ → 44³ | 0.4508 | 0.0606 | 0.0117 |
| `setup_56/setup_56_nuc_128nm_in116_fp32.onnx` | 42.9 | 116³ → 28³ | 0.0004 | 0.0001 | 0.0000 |
| `setup_56/setup_56_nuc_128nm_in116_fp16.onnx` | 21.6 | 116³ → 28³ | 0.0405 | 0.0067 | 0.0046 |
| `setup_56/setup_56_nuc_128nm_in116_int8w.onnx` | 10.9 | 116³ → 28³ | 0.4412 | 0.0668 | 0.0820 |
| `setup_56/setup_56_nuc_128nm_in116_int8w_fp16.onnx` | 10.9 | 116³ → 28³ | 0.3753 | 0.0611 | 0.0638 |
| `setup_55/setup_55_nuc_128nm_in172_fp32.onnx` | 22.5 | 172³ → 84³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_55/setup_55_nuc_128nm_in172_fp16.onnx` | 11.3 | 172³ → 84³ | 0.1344 | 0.0191 | 0.0235 |
| `setup_55/setup_55_nuc_128nm_in172_int8w.onnx` | 5.8 | 172³ → 84³ | 0.5892 | 0.0901 | 0.0995 |
| `setup_55/setup_55_nuc_128nm_in172_int8w_fp16.onnx` | 5.8 | 172³ → 84³ | 0.5302 | 0.0846 | 0.0957 |
| `setup_55/setup_55_nuc_128nm_in140_fp32.onnx` | 22.5 | 140³ → 52³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_55/setup_55_nuc_128nm_in140_fp16.onnx` | 11.3 | 140³ → 52³ | 0.0801 | 0.0183 | 0.0036 |
| `setup_55/setup_55_nuc_128nm_in140_int8w.onnx` | 5.8 | 140³ → 52³ | 0.4925 | 0.0885 | 0.0220 |
| `setup_55/setup_55_nuc_128nm_in140_int8w_fp16.onnx` | 5.8 | 140³ → 52³ | 0.4770 | 0.0798 | 0.0171 |
| `setup_55/setup_55_nuc_128nm_in124_fp32.onnx` | 22.5 | 124³ → 36³ | 0.0000 | 0.0000 | 0.0000 |
| `setup_55/setup_55_nuc_128nm_in124_fp16.onnx` | 11.3 | 124³ → 36³ | 0.0660 | 0.0168 | 0.0021 |
| `setup_55/setup_55_nuc_128nm_in124_int8w.onnx` | 5.8 | 124³ → 36³ | 0.4925 | 0.0864 | 0.0000 |
| `setup_55/setup_55_nuc_128nm_in124_int8w_fp16.onnx` | 5.8 | 124³ → 36³ | 0.4770 | 0.0763 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in148_fp32.onnx` | 42.9 | 148³ → 60³ | 0.0541 | 0.0053 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in148_fp16.onnx` | 21.6 | 148³ → 60³ | 0.0977 | 0.0085 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in148_int8w.onnx` | 10.9 | 148³ → 60³ | 1.2152 | 0.0847 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in148_int8w_fp16.onnx` | 10.9 | 148³ → 60³ | 1.4691 | 0.0912 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in132_fp32.onnx` | 42.9 | 132³ → 44³ | 0.0133 | 0.0015 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in132_fp16.onnx` | 21.6 | 132³ → 44³ | 0.0425 | 0.0061 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in132_int8w.onnx` | 10.9 | 132³ → 44³ | 0.7255 | 0.0742 | 0.0012 |
| `setup_61/setup_61_nuc_256nm_in132_int8w_fp16.onnx` | 10.9 | 132³ → 44³ | 0.8130 | 0.0792 | 0.0012 |
| `setup_61/setup_61_nuc_256nm_in116_fp32.onnx` | 42.9 | 116³ → 28³ | 0.0023 | 0.0003 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in116_fp16.onnx` | 21.6 | 116³ → 28³ | 0.0283 | 0.0052 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in116_int8w.onnx` | 10.9 | 116³ → 28³ | 0.4741 | 0.0666 | 0.0000 |
| `setup_61/setup_61_nuc_256nm_in116_int8w_fp16.onnx` | 10.9 | 116³ → 28³ | 0.4942 | 0.0685 | 0.0000 |
| `setup_67/setup_67_er_8nm_in148_fp32.onnx` | 42.9 | 148³ → 60³ | 0.0004 | 0.0000 | 0.0000 |
| `setup_67/setup_67_er_8nm_in148_fp16.onnx` | 21.6 | 148³ → 60³ | 0.0127 | 0.0017 | 0.0019 |
| `setup_67/setup_67_er_8nm_in148_int8w.onnx` | 10.9 | 148³ → 60³ | 0.1386 | 0.0136 | 0.0171 |
| `setup_67/setup_67_er_8nm_in148_int8w_fp16.onnx` | 10.9 | 148³ → 60³ | 0.1204 | 0.0131 | 0.0153 |
| `setup_67/setup_67_er_8nm_in132_fp32.onnx` | 42.9 | 132³ → 44³ | 0.0003 | 0.0000 | 0.0000 |
| `setup_67/setup_67_er_8nm_in132_fp16.onnx` | 21.6 | 132³ → 44³ | 0.0091 | 0.0020 | 0.0000 |
| `setup_67/setup_67_er_8nm_in132_int8w.onnx` | 10.9 | 132³ → 44³ | 0.0754 | 0.0144 | 0.0047 |
| `setup_67/setup_67_er_8nm_in132_int8w_fp16.onnx` | 10.9 | 132³ → 44³ | 0.0841 | 0.0145 | 0.0012 |
| `setup_67/setup_67_er_8nm_in116_fp32.onnx` | 42.9 | 116³ → 28³ | 0.0001 | 0.0000 | 0.0000 |
| `setup_67/setup_67_er_8nm_in116_fp16.onnx` | 21.6 | 116³ → 28³ | 0.0090 | 0.0022 | 0.0364 |
| `setup_67/setup_67_er_8nm_in116_int8w.onnx` | 10.9 | 116³ → 28³ | 0.0902 | 0.0194 | 0.2870 |
| `setup_67/setup_67_er_8nm_in116_int8w_fp16.onnx` | 10.9 | 116³ → 28³ | 0.0838 | 0.0206 | 0.2824 |

## Finetuning in the browser and PyTorch originals

Each folder's `manifest.json` has a `finetune` block listing the modes the model offers; a model without the block, or
without a mode, is not offered that mode. `source_repo` points to the PyTorch twin (`model_state_dict.pt` + module code
`model_def.py`, plus `model.ts` / `metadata.json` for the cellmap-flow server catalog); 2D students also name their
3D teacher in `distilled_from_repo`.

- **head**: train the final 1×1 conv on frozen features (`features` tensor, `weight` / `bias` initializers).
- **full**: train the whole network with TF.js on WebGPU: `finetune/train_graph.json` (op list, NHWC for the 2D students,
  NDHWC for the 3D models) + the reference runtime `finetune/cellmap_tfjs_train.js`. 2D: parameters named as the ONNX
  initializers (`weight_map` identity). 3D: parameters are the PyTorch state_dict (Conv3d `[O,I,kd,kh,kw]`,
  ConvTranspose3d `[I,O,2,2,2]`), traced from a native-Conv3d export (BatchNorm unfused) and checked by replaying the graph
  in PyTorch against the original model.

`finetune/initializers.json` gives every initializer's byte offset in each `.onnx` (raw_data inline, fixed shapes), so trained
weights are written back in place (as fp16 in the fp16 files) and served as before; int8w files are not writable.
In the 3D folders the served graph is rewritten for onnxruntime-web (Conv3d split into `convs.<k>` Conv2d per depth tap,
ConvTranspose3d k=s=2 as a 1×1 conv to Cout·8 channels + depth-to-space, channel `c·8 + a·4 + b·2 + e`), so:
`finetune/torch_params.json` rebuilds each PyTorch param from the `.onnx` (`from_onnx`), and `finetune/weight_map.json` maps
each ONNX initializer back from its PyTorch param (`to_onnx`). Both are JSON op lists — `index [axis,k]`, `permute`,
`reshape`, `repeat_interleave [axis,n]`, `stack` — e.g. ConvTranspose3d weight `permute [1,2,3,4,0]` + `reshape [Cout·8,Cin,1,1]`,
bias `repeat_interleave [0,8]`; the older Python-syntax `transform` string is kept.

| folder | source_repo | modes | full: step at 556² / 300² (peak) |
|---|---|---|---|
| `setup_54` | [cellmap/mito-aff-resunet-setup-54](https://huggingface.co/cellmap/mito-aff-resunet-setup-54) | head, full | — |
| `setup_53` | [cellmap/mito-aff-unetiso-setup-53](https://huggingface.co/cellmap/mito-aff-unetiso-setup-53) | head, full | — |
| `setup_56` | [cellmap/nuc-resunet-setup-56](https://huggingface.co/cellmap/nuc-resunet-setup-56) | head, full | — |
| `setup_55` | [cellmap/nuc-unetiso-setup-55](https://huggingface.co/cellmap/nuc-unetiso-setup-55) | head, full | — |
| `setup_61` | [cellmap/nuc-resunet-setup-61](https://huggingface.co/cellmap/nuc-resunet-setup-61) | head, full | — |
| `setup_67` | [cellmap/er-resunet-setup-67](https://huggingface.co/cellmap/er-resunet-setup-67) | head, full | — |
| `setup_53_2d` | [cellmap/mito-aff-unetiso-setup-53-2d](https://huggingface.co/cellmap/mito-aff-unetiso-setup-53-2d) | head, full | 1189 ms (3.9 GB) / 308 ms (1.0 GB) |
| `setup_54_2d` | [cellmap/mito-aff-resunet-setup-54-2d](https://huggingface.co/cellmap/mito-aff-resunet-setup-54-2d) | head, full | 1963 ms (10.7 GB) / 490 ms (2.7 GB) |

Full-mode step times (2D students): TF.js 4.22 WebGPU, Chrome, Apple M-series GPU, batch 1, fp32, masked BCE + Adam; peak = TF.js tracked tensors (`manifest.json` → `finetune.full.spike`). No timings published for the 3D models: with plain TF.js autodiff (as in the reference runtime) every intermediate is kept until the backward pass, so only small blocks fit.


## Use in cellmap-flow serverless

ONNX model → remote URL `https://huggingface.co/cellmap/onnx-web-models/resolve/main/<folder>/<file>.onnx`
(e.g. `setup_54/setup_54_mito_affinities_16nm_in148_fp32.onnx`);
Input size `N,N,N`, Output size `N-88,N-88,N-88`, input/output voxel = the model's voxel size, Channels = number of channels
(Has channel axis = true for the 3-channel affinity models). Each folder's `manifest.json` has the full metadata.
Anonymous Hugging Face downloads are rate-limited (HTTP 429), so clients should cache the model bytes.