v3.5: 160px, 277 classes, static width, gated ONNX and CoreML exports
Browse filesReplaces the v2 network on main. Not a drop-in: input height 128 -> 160, output classes 316 -> 277, charset 315 -> 276, and the width axis goes from dynamic to a static 1024 because attention fixes the sequence length at trace time. Revision a51be11 still serves v2 and is unaffected.
Both exports are gated against the PyTorch model on a seeded uniform [-1,1] input: ONNX on logits at rtol=atol=1e-3, CoreML on logits at 1e-3 and on an identical decoded string. The CoreML gate runs CPU-only, so on-device numerics are unverified.
No held-out evaluation exists for this model. The card states what the 0.0210 selection metric is and is not.
- README.md +142 -111
- charset.txt +1 -1
- coreml/monocr.mlpackage.json +16 -0
- coreml/monocr.mlpackage/Data/com.apple.CoreML/model.mlmodel +2 -2
- coreml/monocr.mlpackage/Data/com.apple.CoreML/weights/weight.bin +2 -2
- coreml/monocr.mlpackage/Manifest.json +3 -3
- monocr.json +8 -9
- onnx/charset.txt +1 -1
- onnx/monocr.json +7 -9
- onnx/monocr.onnx +2 -2
- pytorch/monocr.pt +3 -0
README.md
CHANGED
|
@@ -14,131 +14,162 @@ tags:
|
|
| 14 |
- crnn
|
| 15 |
---
|
| 16 |
|
| 17 |
-
# MonOCR
|
| 18 |
|
| 19 |
-
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
-
##
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
|
| 30 |
-
|
|
| 31 |
-
|
|
| 32 |
-
|
|
| 33 |
-
| [monocr-go](https://github.com/MonDevHub/monocr-onnx/tree/main/go) | Go | Go module proxy | v0.1.1 |
|
| 34 |
-
| [monocr-rust](https://github.com/MonDevHub/monocr-onnx/tree/main/rust) | Rust | git dependency | not on crates.io |
|
| 35 |
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
|
| 41 |
-
|
|
|
|
| 42 |
|
| 43 |
-
|
| 44 |
-
| :---------------- | :---------------------- | :------------------------------------------------- |
|
| 45 |
-
| **ONNX** | `onnx/monocr.onnx` | Standard deployments (Server/Desktop). |
|
| 46 |
-
| **Core ML** | `coreml/monocr.mlpackage`| Apple ecosystem (iOS/macOS/watchOS). |
|
| 47 |
-
| **PyTorch** | `pytorch/monocr.ckpt` | Training, fine-tuning, and research. |
|
| 48 |
|
| 49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
-
|
| 61 |
-
split** at epoch 27. The checkpoint published here records it internally
|
| 62 |
-
(`monitor: val_cer`, `best_model_score` 0.0252, step 218,949), so the figure is
|
| 63 |
-
attributable to this artifact. Two things it is not: it is a *validation* number
|
| 64 |
-
rather than a held-out test number, and its split was rendered by the same
|
| 65 |
-
generator **and the same typefaces** as training — font-disjoint splits were
|
| 66 |
-
introduced after this checkpoint was trained. It therefore measures fit to unseen
|
| 67 |
-
*text* in fonts the model had already seen, which is a substantially easier task
|
| 68 |
-
than reading an unseen typeface.
|
| 69 |
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
## Dataset Summary
|
| 77 |
-
|
| 78 |
-
- **Total samples**: 3,030,000
|
| 79 |
-
- **Train size**: 3,000,000
|
| 80 |
-
- **Validation size**: 30,000
|
| 81 |
-
- **Data source description**: Procedural synthetic text generation across multiple Mon fonts combined with real-world digit corpuses.
|
| 82 |
-
- **Augmentation strategy**: Applied during training: image-level augmentations including noise, blur, and transformations.
|
| 83 |
-
|
| 84 |
-
## Model Specifications
|
| 85 |
-
|
| 86 |
-
- **Architecture type**: MobileNetV3-Large Backbone + 2-layer BiLSTM + Linear CTC Head
|
| 87 |
-
- **Parameter count**: 6.58M parameters
|
| 88 |
-
- **Model size**: 100.73 MB (PyTorch Checkpoint)
|
| 89 |
-
- **Training hardware**: NVIDIA GPU (Single GPU run)
|
| 90 |
-
- **Training time**: ~2-4 days
|
| 91 |
-
|
| 92 |
-
## Reproducibility
|
| 93 |
-
|
| 94 |
-
- **Optimizer**: AdamW
|
| 95 |
-
- **Learning rate**: 0.0001 (Warmup + Cosine Annealing)
|
| 96 |
-
- **Batch size**: 48 (with Gradient Accumulation = 4)
|
| 97 |
-
- **Loss function**: CTCLoss (with label smoothing $\epsilon=0.1$)
|
| 98 |
-
|
| 99 |
-
## Technical Specification
|
| 100 |
-
|
| 101 |
-
- **Input Tensors**: Grayscale (1-channel), float32, fixed 128px height. The ONNX graph is `[1, 1, 128, width]`: batch is fixed at 1 and the **width axis is dynamic**.
|
| 102 |
-
- **Image Preprocessing**: Aspect-ratio preserving resize to 128px height, then `pixel / 127.5 - 1.0` into the `[-1, 1]` range. Where padding is used it must be **white** — `+1.0` after normalization, not `0.0`. Training padded to 1024px width; since the width axis is dynamic, 1024 is the width the model was trained at rather than a required input size.
|
| 103 |
-
- **Decoding Strategy**: Connectionist Temporal Classification (CTC) Greedy Decoding (Default).
|
| 104 |
-
- **Vocabulary**: 315 characters (Mon, Burmese, digits, punctuation, and symbols). Encoding is standard UTF-8 (see `charset.txt`).
|
| 105 |
-
|
| 106 |
-
## Integration Guidelines
|
| 107 |
-
|
| 108 |
-
For developers building custom drivers:
|
| 109 |
-
|
| 110 |
-
1. Refer to `charset.txt` for the index-to-character mapping (Index 0 is reserved for `<blank>`).
|
| 111 |
-
2. Ensure input images are high-contrast and properly scaled to 128px height.
|
| 112 |
-
3. The ONNX width axis is dynamic, so crops of differing width can be fed directly without padding to 1024. Batch is fixed at 1.
|
| 113 |
-
4. Read `num_classes` from the JSON sidecar and check it equals `len(charset) + 1` before decoding. If the two disagree, the model will still return confident text — it will just be the wrong text, with no error raised.
|
| 114 |
-
|
| 115 |
-
## Limitations — what no number here covers
|
| 116 |
-
|
| 117 |
-
Stated so that an unmeasured property is not read as a measured one:
|
| 118 |
-
|
| 119 |
-
- **Real photographs.** Every evaluation sample was synthetic, rendered by the
|
| 120 |
-
same generator the model trained on. The reported CER measures generalisation
|
| 121 |
-
to unseen text, not to an unseen renderer. A defect in the renderer is learned,
|
| 122 |
-
validated against and tested against identically, and scores well throughout.
|
| 123 |
-
- **Unseen typefaces.** This checkpoint's validation split shared its fonts with
|
| 124 |
-
training, so no figure here separates typographic generalisation from text
|
| 125 |
-
generalisation.
|
| 126 |
-
- **Page layout and line segmentation.** Inputs are pre-cropped single lines.
|
| 127 |
-
- **Handwriting.** Not covered.
|
| 128 |
-
- **Beam-search decoding.** Greedy only; the reported figures are greedy.
|
| 129 |
-
- **Per-line latency.** No latency tail has been measured.
|
| 130 |
-
|
| 131 |
-
## License
|
| 132 |
-
|
| 133 |
-
All model weights and metadata are provided under the MIT License.
|
| 134 |
|
| 135 |
## Citation
|
| 136 |
|
| 137 |
```bibtex
|
| 138 |
-
@
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
year
|
| 142 |
-
url
|
| 143 |
}
|
| 144 |
```
|
|
|
|
| 14 |
- crnn
|
| 15 |
---
|
| 16 |
|
| 17 |
+
# MonOCR — line-level OCR for Mon (mnw)
|
| 18 |
|
| 19 |
+
A CRNN that reads one line of Mon text and returns a string. Mon is classified by
|
| 20 |
+
UNESCO as a vulnerable language, and this is the model, the charset and the export
|
| 21 |
+
artifacts around it.
|
| 22 |
|
| 23 |
+
**Read the two sections marked "what this number is not" before quoting anything
|
| 24 |
+
from this card.** This model has no held-out evaluation. What it has is a
|
| 25 |
+
training-time selection metric and a like-for-like comparison against the model
|
| 26 |
+
it replaces, and those are different things.
|
| 27 |
|
| 28 |
+
## What changed at v3.5, and what it breaks
|
| 29 |
+
|
| 30 |
+
This revision replaces the v2 network. **It is not a drop-in.** Four parts of the
|
| 31 |
+
contract moved:
|
| 32 |
|
| 33 |
+
| | v2 (revision `a51be11`) | v3.5 (this revision) |
|
| 34 |
+
|---|---|---|
|
| 35 |
+
| Input height | 128 | **160** |
|
| 36 |
+
| Input width | **dynamic** — any width accepted | **static 1024** |
|
| 37 |
+
| Batch axis | fixed at 1 | **dynamic** |
|
| 38 |
+
| Output classes | 316 | **277** |
|
| 39 |
+
| Charset | 315 characters | **276 characters** |
|
| 40 |
+
| Parameters | 6,575,868 | **11,553,437** |
|
|
|
|
|
|
|
| 41 |
|
| 42 |
+
The width axis is the one that catches integrations. v2's graph accepted a crop
|
| 43 |
+
of any width; v3.5 accepts 1024 and nothing else. `nn.MultiheadAttention` fixes
|
| 44 |
+
the sequence length when the graph is traced, so a declaration of anything wider
|
| 45 |
+
would be a promise the artifact cannot keep. **Resize to 1024 before calling.**
|
| 46 |
|
| 47 |
+
If you are pinned to `a51be11` you are unaffected — that revision still serves v2
|
| 48 |
+
and always will. Move deliberately.
|
| 49 |
|
| 50 |
+
## Files
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
+
| Path | What |
|
| 53 |
+
|---|---|
|
| 54 |
+
| `onnx/monocr.onnx` | ONNX FP32, opset 17, 46,247,040 bytes |
|
| 55 |
+
| `onnx/monocr.json` | sidecar: charset, geometry, class count, normalization contract |
|
| 56 |
+
| `onnx/charset.txt` | the 276 characters, index 1..276; **first character is U+0020** |
|
| 57 |
+
| `coreml/monocr.mlpackage` | Core ML, FP32 |
|
| 58 |
+
| `coreml/monocr.mlpackage.json` | its sidecar |
|
| 59 |
+
| `pytorch/monocr.pt` | EMA-merged inference weights, 46,348,409 bytes |
|
| 60 |
+
| `charset.txt`, `monocr.json` | copies at the root, for consumers that expect them there |
|
| 61 |
+
|
| 62 |
+
**Load the charset from the same revision as the weights.** They are one contract:
|
| 63 |
+
a charset from another revision decodes every index to the wrong character and
|
| 64 |
+
raises nothing. Strip only `\n` and `\r` — the first class is a space, and a bare
|
| 65 |
+
`.strip()` removes it and shifts every index by one.
|
| 66 |
+
|
| 67 |
+
## Input contract
|
| 68 |
+
|
| 69 |
+
- Grayscale, 1 channel, float32, shape `[batch, 1, 160, 1024]`
|
| 70 |
+
- Aspect-preserving resize to height 160, then pad the width to 1024
|
| 71 |
+
- Normalize `pixel / 127.5 - 1.0`
|
| 72 |
+
- **Pad with white**, which is `+1.0` after normalization, not `0.0`
|
| 73 |
+
- Greedy CTC decode; index 0 is the blank
|
| 74 |
+
|
| 75 |
+
Feeding raw uint8 in `[0, 255]` produces confident garbage with no error.
|
| 76 |
+
|
| 77 |
+
## Architecture
|
| 78 |
+
|
| 79 |
+
MobileNetV3-Large + squeeze-excitation neck → band pooling → 2×BiLSTM(512) →
|
| 80 |
+
bottleneck self-attention (256-dim, 4 heads) → Linear(1024 → 277) → CTC.
|
| 81 |
+
11,553,437 parameters, all trainable.
|
| 82 |
+
|
| 83 |
+
## Performance, and what these numbers are not
|
| 84 |
+
|
| 85 |
+
### The selection metric
|
| 86 |
+
|
| 87 |
+
`val_cer` **0.0210**, greedy, at epoch 23 / step 93,645.
|
| 88 |
+
|
| 89 |
+
**What it is not:** an evaluation. Four things travel with it or it says something
|
| 90 |
+
it does not mean.
|
| 91 |
+
|
| 92 |
+
1. **4,096 lines, not the validation split.** It is a periodic check during
|
| 93 |
+
training, not a pass over the data.
|
| 94 |
+
2. **One typeface.** Every one of those lines is rendered in a single font.
|
| 95 |
+
3. **It selected the checkpoint.** The same number drove `save_top_k`, so it
|
| 96 |
+
cannot also be an unbiased estimate of that checkpoint.
|
| 97 |
+
4. **Synthetic.** Rendered by the same generator the model trained on.
|
| 98 |
+
|
| 99 |
+
**There is no held-out evaluation of this model.** The test split is unspent. No
|
| 100 |
+
number on this card describes accuracy on a photographed page.
|
| 101 |
+
|
| 102 |
+
### Against v2, on identical images
|
| 103 |
+
|
| 104 |
+
The one comparison that is available and fair. Both generations ship an ONNX
|
| 105 |
+
export, so the same rendered lines go through both graphs with the same
|
| 106 |
+
preprocessing and the same greedy decode. Text is restricted to the **273
|
| 107 |
+
characters both charsets can emit**, so neither model is charged for a character
|
| 108 |
+
it has no class for. 600 lines per arm, rendered 2026-08-15.
|
| 109 |
+
|
| 110 |
+
| Rendered in | n | v2 CER | v3.5 CER | error reduced |
|
| 111 |
+
|---|---:|---:|---:|---:|
|
| 112 |
+
| the 32 trained designs | 600 | 0.1470 | **0.0396** | 73% |
|
| 113 |
+
| `namkhon`, held out from v3.5 | 600 | 0.0521 | **0.0188** | 64% |
|
| 114 |
+
| `pyidaungsu`, `yunghkio`, held out from v3.5 | 600 | 0.0342 | **0.0051** | 85% |
|
| 115 |
+
|
| 116 |
+
The bottom two rows are held out from v3.5 and from nothing for v2 — v2 predates
|
| 117 |
+
font-disjoint splits — so the comparison is tilted against v3.5 and it wins
|
| 118 |
+
anyway. **This is a preview, not an evaluation:** one rendering pipeline reading
|
| 119 |
+
its own output, synthetic for both models. It supports a narrow claim, that v3.5
|
| 120 |
+
reads rendered Mon better than v2, and no broader one.
|
| 121 |
+
|
| 122 |
+
### Reading a page
|
| 123 |
+
|
| 124 |
+
Wide lines need care. Measured 2026-08-15 over 240 rendered lines wide enough to
|
| 125 |
+
need the choice, median 3 model windows each: squeezing a whole line into the
|
| 126 |
+
1024px canvas scored **0.1434**, cutting it into canvas-width tiles at whitespace
|
| 127 |
+
columns and joining scored **0.0795**. **Tile wide lines for this model.** The
|
| 128 |
+
opposite is true of v2, where squeezing wins — the direction is a property of the
|
| 129 |
+
network, so re-measure if you swap models.
|
| 130 |
+
|
| 131 |
+
## Exports, and how they were checked
|
| 132 |
+
|
| 133 |
+
Both exports are gated against the PyTorch model on a seeded uniform `[-1, 1]`
|
| 134 |
+
input, and the gate fails the build:
|
| 135 |
+
|
| 136 |
+
- **ONNX** — logits within `rtol=atol=1e-3`.
|
| 137 |
+
- **Core ML** — logits within `1e-3` **and** an identical CTC-decoded string.
|
| 138 |
+
|
| 139 |
+
The Core ML gate runs on CPU. The Neural Engine computes in fp16 and its compiler
|
| 140 |
+
may reassociate, so what ships to a device is not what was verified. **No device
|
| 141 |
+
latency number exists for this model**, on any platform.
|
| 142 |
+
|
| 143 |
+
## Limitations
|
| 144 |
|
| 145 |
+
- **Real photographs.** Every training sample is synthetic, rendered by one
|
| 146 |
+
generator. A camera photo of physical text is out of domain and fails
|
| 147 |
+
confidently — measured on a whiteboard photo, the model returned fluent Mon at
|
| 148 |
+
confidence 0.83 for text that appears nowhere on the page. **Confidence is not
|
| 149 |
+
a usable filter for this.**
|
| 150 |
+
- **Page layout.** The input is a cropped line. There is no detector and no
|
| 151 |
+
reading-order model.
|
| 152 |
+
- **Handwriting**, except Myanmar digits.
|
| 153 |
+
- **Beam search.** Greedy only in the exported graph.
|
| 154 |
+
- **Latency.** No per-line tail measured, and no on-device number at all.
|
| 155 |
+
- **Typeface coverage.** 35 distinct designs across 79 usable font files bound
|
| 156 |
+
every generalisation claim above.
|
| 157 |
|
| 158 |
+
## Training data
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 159 |
|
| 160 |
+
Rendered from a Mon text corpus. **The corpus licensing is not fully settled**:
|
| 161 |
+
one source has no established licence, and CC BY-SA attribution for another
|
| 162 |
+
cannot be reconstructed because per-article URLs were stripped at import. The MIT
|
| 163 |
+
licence on this repository covers the model weights and the code that produced
|
| 164 |
+
them; it does not resolve the provenance of the text they were rendered from.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## Citation
|
| 167 |
|
| 168 |
```bibtex
|
| 169 |
+
@misc{monocr,
|
| 170 |
+
title = {MonOCR: line-level OCR for the Mon language},
|
| 171 |
+
author = {Zin Min},
|
| 172 |
+
year = {2026},
|
| 173 |
+
url = {https://huggingface.co/janakhpon/monocr}
|
| 174 |
}
|
| 175 |
```
|
charset.txt
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
!"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒː
|
|
|
|
| 1 |
+
!"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¼½¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒːπကခဂဃငစဆဇဈဉညဋဌဍဎဏတထဒဓနပဖဗဘမယရလဝသဟဠအဣဤဥဦဧဨဩဪါာိီုူေဲဳဴဵံ့း္်ျြွှဿ၀၁၂၃၄၅၆၇၈၉၊။၌၍၎၏ၐၑၓၚၛၜၝၞၟၠ–—‘’‚“”•…−
|
coreml/monocr.mlpackage.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"charset": " !\"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¼½¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒːπကခဂဃငစဆဇဈဉညဋဌဍဎဏတထဒဓနပဖဗဘမယရလဝသဟဠအဣဤဥဦဧဨဩဪါာိီုူေဲဳဴဵံ့း္်ျြွှဿ၀၁၂၃၄၅၆၇၈၉၊။၌၍၎၏ၐၑၓၚၛၜၝၞၟၠ–—‘’‚“”•…−",
|
| 3 |
+
"img_height": 160,
|
| 4 |
+
"img_width": 1024,
|
| 5 |
+
"model_version": "3.5",
|
| 6 |
+
"format": "mlpackage",
|
| 7 |
+
"architecture": "MobileNetV3-Large + SE + 2xBiLSTM(512) + BottleneckAttention + CTC",
|
| 8 |
+
"num_classes": 277,
|
| 9 |
+
"rnn_hidden_size": 512,
|
| 10 |
+
"rnn_layers": 2,
|
| 11 |
+
"backbone": "mobilenetv3_large",
|
| 12 |
+
"input_normalization": "pixel / 127.5 - 1.0",
|
| 13 |
+
"input_dtype": "float32",
|
| 14 |
+
"input_channels": 1,
|
| 15 |
+
"input_colorspace": "grayscale"
|
| 16 |
+
}
|
coreml/monocr.mlpackage/Data/com.apple.CoreML/model.mlmodel
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b8ad3699faf4a4644812c58867ca214997e956f1522a485804c97cbe065fca11
|
| 3 |
+
size 97599
|
coreml/monocr.mlpackage/Data/com.apple.CoreML/weights/weight.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aaa035c39c34753c2695e3e25766eb3d71d8bf291536244d48b4857206834698
|
| 3 |
+
size 46176000
|
coreml/monocr.mlpackage/Manifest.json
CHANGED
|
@@ -1,18 +1,18 @@
|
|
| 1 |
{
|
| 2 |
"fileFormatVersion": "1.0.0",
|
| 3 |
"itemInfoEntries": {
|
| 4 |
-
"
|
| 5 |
"author": "com.apple.CoreML",
|
| 6 |
"description": "CoreML Model Weights",
|
| 7 |
"name": "weights",
|
| 8 |
"path": "com.apple.CoreML/weights"
|
| 9 |
},
|
| 10 |
-
"
|
| 11 |
"author": "com.apple.CoreML",
|
| 12 |
"description": "CoreML Model Specification",
|
| 13 |
"name": "model.mlmodel",
|
| 14 |
"path": "com.apple.CoreML/model.mlmodel"
|
| 15 |
}
|
| 16 |
},
|
| 17 |
-
"rootModelIdentifier": "
|
| 18 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"fileFormatVersion": "1.0.0",
|
| 3 |
"itemInfoEntries": {
|
| 4 |
+
"4B1BEA03-F6E1-4E73-9263-1A15DF6E48E2": {
|
| 5 |
"author": "com.apple.CoreML",
|
| 6 |
"description": "CoreML Model Weights",
|
| 7 |
"name": "weights",
|
| 8 |
"path": "com.apple.CoreML/weights"
|
| 9 |
},
|
| 10 |
+
"E70682BC-0A4A-45B0-B67C-83338031F87D": {
|
| 11 |
"author": "com.apple.CoreML",
|
| 12 |
"description": "CoreML Model Specification",
|
| 13 |
"name": "model.mlmodel",
|
| 14 |
"path": "com.apple.CoreML/model.mlmodel"
|
| 15 |
}
|
| 16 |
},
|
| 17 |
+
"rootModelIdentifier": "E70682BC-0A4A-45B0-B67C-83338031F87D"
|
| 18 |
}
|
monocr.json
CHANGED
|
@@ -1,17 +1,16 @@
|
|
| 1 |
{
|
| 2 |
-
"charset": " !\"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒː
|
| 3 |
-
"img_height":
|
| 4 |
"img_width": 1024,
|
| 5 |
-
"model_version": "
|
| 6 |
-
"
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
"rnn_hidden_size": 384,
|
| 10 |
"rnn_layers": 2,
|
| 11 |
"backbone": "mobilenetv3_large",
|
| 12 |
"input_normalization": "pixel / 127.5 - 1.0",
|
| 13 |
"input_dtype": "float32",
|
| 14 |
"input_channels": 1,
|
| 15 |
"input_colorspace": "grayscale",
|
| 16 |
-
"
|
| 17 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"charset": " !\"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¼½¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒːπကခဂဃငစဆဇဈဉညဋဌဍဎဏတထဒဓနပဖဗဘမယရလဝသဟဠအဣဤဥဦဧဨဩဪါာိီုူေဲဳဴဵံ့း္်ျြွှဿ၀၁၂၃၄၅၆၇၈၉၊။၌၍၎၏ၐၑၓၚၛၜၝၞၟၠ–—‘’‚“”•…−",
|
| 3 |
+
"img_height": 160,
|
| 4 |
"img_width": 1024,
|
| 5 |
+
"model_version": "3.5",
|
| 6 |
+
"architecture": "MobileNetV3-Large + SE + BiLSTM + BottleneckAttention + CTC",
|
| 7 |
+
"num_classes": 277,
|
| 8 |
+
"rnn_hidden_size": 512,
|
|
|
|
| 9 |
"rnn_layers": 2,
|
| 10 |
"backbone": "mobilenetv3_large",
|
| 11 |
"input_normalization": "pixel / 127.5 - 1.0",
|
| 12 |
"input_dtype": "float32",
|
| 13 |
"input_channels": 1,
|
| 14 |
"input_colorspace": "grayscale",
|
| 15 |
+
"format": "pytorch"
|
| 16 |
+
}
|
onnx/charset.txt
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
!"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒː
|
|
|
|
| 1 |
+
!"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¼½¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒːπကခဂဃငစဆဇဈဉညဋဌဍဎဏတထဒဓနပဖဗဘမယရလဝသဟဠအဣဤဥဦဧဨဩဪါာိီုူေဲဳဴဵံ့း္်ျြွှဿ၀၁၂၃၄၅၆၇၈၉၊။၌၍၎၏ၐၑၓၚၛၜၝၞၟၠ–—‘’‚“”•…−
|
onnx/monocr.json
CHANGED
|
@@ -1,18 +1,16 @@
|
|
| 1 |
{
|
| 2 |
-
"charset": " !\"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒː
|
| 3 |
-
"img_height":
|
| 4 |
"img_width": 1024,
|
| 5 |
"opset_version": 17,
|
| 6 |
-
"model_version": "
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
"
|
| 10 |
-
"rnn_hidden_size": 384,
|
| 11 |
"rnn_layers": 2,
|
| 12 |
"backbone": "mobilenetv3_large",
|
| 13 |
"input_normalization": "pixel / 127.5 - 1.0",
|
| 14 |
"input_dtype": "float32",
|
| 15 |
"input_channels": 1,
|
| 16 |
-
"input_colorspace": "grayscale"
|
| 17 |
-
"_provenance": "Charset repaired 2026-08-14. This sidecar declared a 912-character corpus-scanned charset against a 316-class model, so 150 of the 315 decodable indices mapped to the wrong character and any consumer decoded Mon as Latin/IPA text with no error and no lookup miss. The charset here is the 315-character allowlist read from charset.txt in this repository (git blob 08b1c23299dd1f68f14b7bdb4be586dd52d27c5d); it is byte-identical to the repaired sidecar in mon_OCR data/models/archive-v2-2026-03/, which was confirmed by decoding validation samples against their labels. num_classes is recorded here so this class of mismatch is checkable: CTC reserves index 0 for blank, so num_classes must equal len(charset) + 1. It is confirmed by fc.weight (316, 768) and hyper_parameters['num_classes'] in the published monocr.ckpt. The 912-character string arrived in commit df456df, a Core ML export that overwrote both this file and the root monocr.json with the Core ML variant -- which is also why the ONNX sidecar had lost opset_version and declared format 'mlpackage'. A re-export that writes only one sidecar will reintroduce that. model_version stays '2.0': mon_OCR names this artifact '2.0-legacy' internally, but that string is a judgement relative to unpublished v3.5 code and changing a published contract field could break an exact-match check in the four SDKs the model card lists. The v2 graph takes [1, 1, 128, width] float32 grayscale: batch is fixed at 1 and the width axis is dynamic. img_width 1024 is the width training preprocessing pads to, not a graph constraint -- README.md states both and reads as a contradiction. opset_version 17 is read from the v2 fp32, fp16 and int8 ONNX exports, the fp32 of which is byte-identical to the ONNX published here at commit d29385b; the binary currently published (LFS oid f212ab7e) is a later re-export that has not been re-inspected."
|
| 18 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"charset": " !\"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~£¥¦§©«¬°±²³´µ·¸¹º»¼½¾ÀÁÂÄÅÆÇÉÊÌÍÑÓÖרÜÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüþĀāıŒœŠšŽžƒːπကခဂဃငစဆဇဈဉညဋဌဍဎဏတထဒဓနပဖဗဘမယရလဝသဟဠအဣဤဥဦဧဨဩဪါာိီုူေဲဳဴဵံ့း္်ျြွှဿ၀၁၂၃၄၅၆၇၈၉၊။၌၍၎၏ၐၑၓၚၛၜၝၞၟၠ–—‘’‚“”•…−",
|
| 3 |
+
"img_height": 160,
|
| 4 |
"img_width": 1024,
|
| 5 |
"opset_version": 17,
|
| 6 |
+
"model_version": "3.5",
|
| 7 |
+
"architecture": "MobileNetV3-Large + SE + BiLSTM + BottleneckAttention + CTC",
|
| 8 |
+
"num_classes": 277,
|
| 9 |
+
"rnn_hidden_size": 512,
|
|
|
|
| 10 |
"rnn_layers": 2,
|
| 11 |
"backbone": "mobilenetv3_large",
|
| 12 |
"input_normalization": "pixel / 127.5 - 1.0",
|
| 13 |
"input_dtype": "float32",
|
| 14 |
"input_channels": 1,
|
| 15 |
+
"input_colorspace": "grayscale"
|
|
|
|
| 16 |
}
|
onnx/monocr.onnx
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b95de1ea0e3dc99a5c31bea32e220835da801f683ed17091eaec9954c80a4c04
|
| 3 |
+
size 46247040
|
pytorch/monocr.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c779f69016ca87da570a474d4f38f8c6110502540485d428f8f90d6d8e03d44e
|
| 3 |
+
size 46348409
|