File size: 1,710 Bytes
2ea4e1d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
---

library_name: onnxruntime
pipeline_tag: object-detection
license: mit
base_model: microsoft/table-transformer-detection
tags:
  - table-detection
  - table-structure-recognition
  - tatr
  - table-transformer
  - onnx
  - gmft
---


# Microsoft Table Transformer (TATR) — ONNX re-export

Re-export of Microsoft's Table Transformer to ONNX, packaged the way [GMFT](https://github.com/conjuncts/gmft) consumes it. Used by [ScanIndex](https://github.com/welcomyou/scanindex) for table detection + structure recognition during DOCX export.

## Variants

| Subdir | Upstream | Task |
|---|---|---|
| `gmft_onnx/detection/model.onnx` | [`microsoft/table-transformer-detection`](https://huggingface.co/microsoft/table-transformer-detection) | Detect table bounding boxes on a page |
| `gmft_onnx/structure/model.onnx` | [`microsoft/table-transformer-structure-recognition-v1.1-all`](https://huggingface.co/microsoft/table-transformer-structure-recognition-v1.1-all) | Detect rows / columns / cells inside a cropped table |

Each subdir also contains the HF `config.json` + preprocessor metadata so `transformers` / `optimum` can wrap the ONNX directly.

## Loading

```python

from huggingface_hub import snapshot_download

local = snapshot_download("welcomyou/gmft-tatr-onnx", local_dir="models")

# Detection:  f"{local}/gmft_onnx/detection/model.onnx"

# Structure:  f"{local}/gmft_onnx/structure/model.onnx"

```

## Re-export reproduction

See [train-convert/gmft/convert/export_gmft_tatr_to_onnx.py](https://github.com/welcomyou/scanindex/blob/main/train-convert/gmft/convert/export_gmft_tatr_to_onnx.py).

## License

MIT, inherited from Microsoft Table Transformer.