"beit": {

"model_src_dir": "image_recognition/beit/quantization/ptq_static",

"dataset_location": "/tf_dataset/pytorch/ImageNet/raw",

"input_model": "/tf_dataset2/models/onnx/beit/beit_base_patch16_224_pt22k_ft22kto1k.onnx",

"main_script": "main.py",

"batch_size": 1

},

<tr>

<td>BEiT</td>

<td>Image Recognition</td>

<td>Post-Training Static Quantization</td>

<td><a href="./onnxrt/image_recognition/beit/quantization/ptq_static">qlinearops</a></td>

</tr>

Step-by-Step

This example load [BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254)(BEiT) model and confirm its accuracy and performance based on [ImageNet-1k dataset](https://www.image-net.org/). You need to download this dataset yourself.

In this example, the BEiT model is pre-trained in a self-supervised fashion on ImageNet-22k - also called ImageNet-21k (14 million images, 21,841 classes) at resolution 224x224, and fine-tuned on the same dataset at resolution 224x224. It was first released in [this repository](https://github.com/microsoft/unilm/tree/master/beit).

pip install neural-compressor

pip install -r requirements.txt

Prepare DETR R18 model for table structure recognition.

python prepare_model.py --input_model=beit_base_patch16_224 --output_model=beit_base_patch16_224_pt22k_ft22kto1k.onnx

Download and extract [ImageNet-1k](https://www.image-net.org/) to dir: /path/to/imagenet. The dir include below folder:

ls /path/to/imagenet

train val

Quantize model with QLinearOps:

bash run_quant.sh --input_model=/path/to/model \ # model path as *.onnx

--dataset_location=/path/to/imagenet \

--output_model=/path/to/save \

--quant_format="QOperator"

bash run_benchmark.sh --input_model=/path/to/model \ # model path as *.onnx

--dataset_location=/path/to/imagenet \

--batch_size=batch_size \

--mode=performance # or accuracy