| --- |
| license: mit |
| widget: |
| - src: https://www.invoicesimple.com/wp-content/uploads/2018/06/Sample-Invoice-printable.png |
| example_title: Invoice |
| --- |
| |
| # Table Transformer (fine-tuned for Table Detection) |
|
|
| Table Transformer (DETR) model trained on PubTables1M. It was introduced in the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Smock et al. and first released in [this repository](https://github.com/microsoft/table-transformer). |
|
|
| Disclaimer: The team releasing Table Transformer did not write a model card for this model so this model card has been written by the Hugging Face team. |
|
|
| ## Model description |
|
|
| The Table Transformer is equivalent to [DETR](https://huggingface.co/docs/transformers/model_doc/detr), a Transformer-based object detection model. Note that the authors decided to use the "normalize before" setting of DETR, which means that layernorm is applied before self- and cross-attention. |
|
|
| ## Usage |
|
|
| You can use the raw model for detecting tables in documents. See the [documentation](https://huggingface.co/docs/transformers/main/en/model_doc/table-transformer) for more info. |
|
|
| ### Run as ONNX (CPU / NPU / GPU) |
|
|
| Detect tables ~14× faster on a Windows NPU at half the model size, with mAP within 1% of the original PyTorch checkpoint — by exporting this model to ONNX. You can also export to ONNX to run on CPU or GPU. |
|
|
| Benchmarked on an Intel Core Ultra 7 258V (PubTables-1M validation, 1000 samples): |
|
|
| | Model | Device | Precision | mAP | mean latency (ms) | p50 latency (ms) | Size (MB) | |
| |---------|--------------|-------------|--------|-------------------|------------------|-----------| |
| | PyTorch | CPU | fp32 | 0.9887 | 620.9 | 600.3 | 115 | |
| | ONNX | OpenVINO NPU | w8a16 (QDQ) | 0.9822 | 44.1 | 41.6 | 58 | |
|
|
| - **How to convert** — Export and quantize with [Microsoft's WinML CLI](https://github.com/microsoft/winml-cli). The NPU build is QDQ-quantized to w8a16; fp32 builds for CPU and GPU are also supported. End-to-end build, evaluation, and a Python inference example: [examples/microsoft-table-transformer-detection](https://github.com/microsoft/winml-cli/blob/main/examples/microsoft-table-transformer-detection/README.md). |
| - **How to run on Windows** — Use [Windows ML](https://learn.microsoft.com/en-us/windows/ai/new-windows-ml/overview), which manages execution providers for NPU / GPU / CPU and routes ONNX inference to the right backend automatically. |
| - **How to run on other platforms** — Use [ONNX Runtime](https://onnxruntime.ai/docs/) with the execution provider of your choice (OpenVINO, QNN, DirectML, CUDA, CPU, etc.). |
|
|
