YOLOX-PAI Turtorial¶

Introduction¶

Welcome to YOLOX-PAI! YOLOX-PAI is an incremental work of YOLOX based on PAI-EasyCV. We use various existing detection methods and PAI-Blade to boost the performance. We also provide an efficient way for end2end object detction.

In breif, our main contributions are:

Investigate various detection methods upon YOLOX to achieve SOTA object detection results.
Provide an easy way to use PAI-Blade to accelerate the inference process.
Provide a convenient way to train/evaluate/export YOLOX-PAI model and conduct end2end object detection.

To learn more details of YOLOX-PAI, you can refer to our technical report or arxiv paper.

Data preparation¶

To download the dataset, please refer to prepare_data.md.

Yolox support both coco format and PAI-Itag detection format,

COCO format¶

To use coco data to train detection, you can refer to 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 train detection, you can refer to configs/detection/yolox/yolox_s_8xb16_300e_coco_pai.py for more configuration details.

Quick Start¶

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

You can use the quick_start.md for local installation or use our provided doker images (for both training and inference).

Pull Docker¶

sudo docker pull registry.cn-shanghai.aliyuncs.com/pai-ai-test/pai-easycv:yolox-pai

Start Container¶

sudo nvidia-docker run -it -v path:path --name easycv_yolox_pai --shm-size=10g --network=host registry.cn-shanghai.aliyuncs.com/pai-ai-test/pai-easycv:yolox-pai

Train¶

Single gpu:

python tools/train.py \
		${CONFIG_PATH} \
		--work_dir ${WORK_DIR}

Multi gpus:

bash tools/dist_train.sh \
		${NUM_GPUS} \
		${CONFIG_PATH} \
		--work_dir ${WORK_DIR}

Arguments

NUM_GPUS: number of gpus
CONFIG_PATH: the config file path of a detection method
WORK_DIR: your path to save models and logs

Examples:

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

GPUS=8
bash tools/dist_train.sh configs/detection/yolox/yolox_s_8xb16_300e_coco.py $GPUS

Evaluation¶

The pretrained model of YOLOX-PAI can be found here.

Single gpu:

python tools/eval.py \
		${CONFIG_PATH} \
		${CHECKPOINT} \
		--eval

Multi gpus:

bash tools/dist_test.sh \
		${CONFIG_PATH} \
		${NUM_GPUS} \
		${CHECKPOINT} \
		--eval

Arguments

CONFIG_PATH: the config file path of a detection method
NUM_GPUS: number of gpus
CHECKPOINT: the checkpoint file named as epoch_*.pth.

Examples:

GPUS=8
bash tools/dist_test.sh configs/detection/yolox/yolox_s_8xb16_300e_coco.py $GPUS work_dirs/detection/yolox/epoch_300.pth --eval

Export model¶

python tools/export.py \
		${CONFIG_PATH} \
		${CHECKPOINT} \
		${EXPORT_PATH}

For more details of the export process, you can refer to export.md.

Arguments

CONFIG_PATH: the config file path of a detection method
CHECKPOINT:your checkpoint file of a detection method named as epoch_*.pth.
EXPORT_PATH: your path to save export model

Examples:

python tools/export.py configs/detection/yolox/yolox_s_8xb16_300e_coco.py \
        work_dirs/detection/yolox/epoch_300.pth \
        work_dirs/detection/yolox/epoch_300_export.pth

Inference¶

Download exported models(preprocess, model, meta) or export your own model. Put them in the following format:

export_blade/
epoch_300_pre_notrt.pt.blade
epoch_300_pre_notrt.pt.blade.config.json
epoch_300_pre_notrt.pt.preprocess

Download test_image

import cv2
from easycv.predictors import TorchYoloXPredictor

output_ckpt = 'export_blade/epoch_300_pre_notrt.pt.blade'
detector = TorchYoloXPredictor(output_ckpt,use_trt_efficientnms=False)

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

# visualize image
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)

cv2.imwrite('result.jpg',image)