EasyCV/docs/source/tutorials/compression.md

# YOLOX Compression Tutorial

## ENV preparation
### install torch
blade_compression needs torch >= 1.10.0
```shell
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=10.2 -c pytorch
```
### install mmcv
with torch >=1.10.0, mmcv >=1.3.16 is required
```shell
pip install mmcv-full==1.4.4 -f https://download.openmmlab.com/mmcv/dist/{cu_version}/{torch_version}/index.html
```
### install onnx and pai_nni
```shell
pip install onnx
pip install https://pai-nni.oss-cn-zhangjiakou.aliyuncs.com/release/2.5/pai_nni-2.5-py3-none-manylinux1_x86_64.whl
```
### install blade_compression
```shell
pip install http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/third_party/blade_compression-0.0.1-py3-none-any.whl
```

## Data preparation
To download the dataset, please refer to [prepare_data.md](../prepare_data.md).

The data is used to train and test compression model. So the data format should be the same of train data format.

### COCO format
To use coco data to eval model, you can refer to [configs/detection/yolox/yolox_s_8xb16_300e_coco.py](../../configs/detection/yolox/yolox_s_8xb16_300e_coco.py) for more configuration details.

### PAI-Itag detection format
To use pai-itag detection format data to eval detection, you can refer to [configs/detection/yolox/yolox_s_8xb16_300e_coco_pai.py](../../configs/detection/yolox/yolox_s_8xb16_300e_coco_pai.py) for more configuration details.

## Local & PAI-DSW

To use COCO format data, use config file `configs/detection/yolox/yolox_s_8xb16_300e_coco.py`

To use PAI-Itag format data, use config file `configs/detection/yolox/yolox_s_8xb16_300e_coco_pai.py`


### Compression
**Quantize:**
This is used to quantize yolox model; The quantized model will be saved in work_dir.

```shell
python tools/quantize.py \
		${CONFIG_PATH} \
		${MODEL_PATH} \
		--work_dir ${WORK_DIR} \
		--device ${DEVICE} \
		--backend ${BACKEND}
```


<details>
<summary>Arguments</summary>

- `CONFIG_PATH`: the config file path of a detection method

- `WORK_DIR`: your path to save models and logs

- `MODEL_PATH`: the models to be quantized

- `DEVICE`: the device quantized models use (cpu/arm)

- `BACKEND`: the quantized models's framework (PyTorch/MNN)

</details>

**Examples:**

Edit `data_root`path in the `${CONFIG_PATH}` to your own data path.

```shell
python tools/quantize.py \
		configs/detection/yolox/yolox_s_8xb16_300e_coco.py \
		models/yolox_s.pth \
		--device cpu \
		--backend PyTorch
```

**Prune:**
This is used to prune yolox model; The pruned model will be saved in work_dir.

```shell
python tools/prune.py \
		${CONFIG_PATH} \
		${MODEL_PATH} \
		--work_dir ${WORK_DIR} \
		--pruning_class ${PRUNING_CLASS} \
		--pruning_algorithm ${PRUNING_ALGORITHM}
```


<details>
<summary>Arguments</summary>

- `CONFIG_PATH`: the config file path of a detection method

- `WORK_DIR`: your path to save models and logs

- `MODEL_PATH`: the quantized models

- `PRUNING_CLASS`: pruning class for pruning models (AGP)

- `PRUNING_ALGORITHM`: pruning algorithm using by pruning class (taylorfo)
</details>

**Examples:**

Edit `data_root`path in the `${CONFIG_PATH}` to your own data path.

```shell
python tools/prune.py \
		configs/detection/yolox/yolox_s_8xb16_300e_coco.py \
		models/yolox_s.pth \
		--pruning_class AGP \
		--pruning_algorithm taylorfo
```

### Evaluation

Model will be auto-eval after compressing. So the result will be writed in the log.

### Inference
Download [test_image](http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/data/small_coco_demo/val2017/000000017627.jpg)

```python
import cv2
from easycv.predictors import TorchYoloXPredictor

output_ckpt = 'work_dirs/compression/yolox/quantize_model.pt'
detector = TorchYoloXPredictor(output_ckpt)

img = cv2.imread('000000017627.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
output = detector.predict([img])
print(output)

# visualize image
from matplotlib import pyplot as plt
image = img.copy()
for box, cls_name in zip(output[0]['detection_boxes'], output[0]['detection_class_names']):
    # box is [x1,y1,x2,y2]
    box = [int(b) for b in box]
    image = cv2.rectangle(image, tuple(box[:2]), tuple(box[2:4]), (0,255,0), 2)
    cv2.putText(image, cls_name, (box[0], box[1]-5), cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0,0,255), 2)
plt.imshow(image)
plt.show()
```