# YOLOX: Exceeding YOLO Series in 2021
> [name=JayHsu][time=Thu, Nov 18, 2021 2:19 PM]

[TOC]
## Install

git clone [e-mail]:Megvii-BaseDetection/YOLOX.git
cd YOLOX
pip3 install -U pip && pip3 install -r requirements.txt
python3 setup.py develop

## Inference
1. 把install時下載的YOLOX/yolox資料夾copy至專案資料夾

scp -r YOLOX/yolox  face_detect/

2. Import相關package

import torch
from yolox.data.data_augment import ValTransform
from yolox.data.datasets import COCO_CLASSES
from yolox.utils import postprocess as yolox_postprocess

3. Load model

exp = yolox_s_Exp()
model = exp.get_model()
model.cuda();
model.eval();
ckpt_file = 'pretrain/yolox_s.pth'
ckpt = torch.load(ckpt_file, map_location="cpu")
model.load_state_dict(ckpt["model"])

4. Inference

img = img.cuda()
outputs = self.model(img)
outputs = yolox_postprocess(
    outputs, self.num_classes, self.confthre,
    self.nmsthre, class_agnostic=True)

5. Sample Code:
http://10.109.6.14:9999/notebooks/Jay/face_detect/yoloX_Inference.ipynb

## Training on custom data
1. 準備dataset, 必須是COCO format
資料夾必須符合固定格式
- dataset folder
- annotation folder
- train_label.json
- val_label.json
- train img folder
- val img folder


2. 準備設定檔(EXP file)
根據選擇的model, 使用對應的模板修改
- data_dir為dataset folder位置
- train_ann: tr annotations (必須放在annotation folder內)
- val_ann: val annotations (必須放在annotation folder內)
- tr_img_folder: tr images
- val_img_folder: val images
- self.num_classes: number of classes

self.data_dir = '/mnt/hdd1/Data/YOLOX/coco128'
self.train_ann = "instances_train2017.json"
self.val_ann = "instances_val2017.json"
self.tr_img_folder = 'images'
self.val_img_folder = 'images'

設定檔模板: http://10.109.6.14:9999/edit/Jay/YOLOX/exps/test/yolox_s.py

3. Train

#cd /home/tpe-aa-02/AA/Jay/YOLOX
cd /mnt/hdd1/AA/Jay/YOLOX

export PYTHONPATH="${PYTHONPATH}:/mnt/hdd1/AA/Jay/YOLOX"


python3 tools/train.py -f exps/test/yolox_s.py -d 1 -b 32 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_s.pth

#start from latest checkpoint
python3 tools/train.py -f exps/test/yolox_s.py -d 1 -b 32 --fp16 -o -c /home/tpe-aa-02/AA/Jay/YOLOX/YOLOX_outputs/yolox_s/last_epoch_ckpt_tmp.pth

#TWCC
python3 tools/train.py -f exps/test/yolox_m.py -d 1 -b 48 --fp16 -o -c pretrain/yolox_m.pth

#DCT 
python3 tools/train.py -f exps/test/yolox_s_dct.py -d 1 -b 32 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_s.pth
python3 tools/train.py -f exps/test/yolox_tiny_dct.py -d 1 -b 32 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_tiny.pth
python3 tools/train.py -f exps/test/nano_dct.py -d 1 -b 32 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_nano.pth
python3 tools/train.py -f exps/test/yolox_m_dct.py -d 1 -b 32 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_m.pth

#IO Location
python3 tools/train.py -f exps/test/yolox_s_ioL.py -d 1 -b 8 --fp16 -o -c /mnt/hdd1/Model/YOLOX/pretrain/yolox_s.pth

#Flower
python3 tools/train.py -f exps/test/yolox_s_flower.py -d 1 -b 32 --fp16 -o -c /home/tpe-aa-02/AA/Jay/YOLOX/YOLOX_outputs/yolox_s_flower/best_ckpt.pth

-f: 設定檔位置
-c: pretarin model位置
-d: 用幾張GPU
-b: batch size
-0: occupy GPU memory first for training

4. training後的model儲存位置

YOLOX/YOLOX_outputs

## Appendix A - YOLOX Performance
- 官方數據

## Appendix B - Error Fix
**1. numpy.float64' object cannot be interpreted as an integer**

#training時出現以下error
TypeError: 'numpy.float64' object cannot be interpreted as an integer

參考issue解決方法: https://github.com/Megvii-BaseDetection/YOLOX/issues/435
./.local/lib/python3.6/site-packages/pycocotools/cocoeval.py

**2. out of memory**
在training中修改以下設定 -d 1 -b 8
-d 1: 只用一張GPU
-b 8: batch size 8

**3. ModuleNotFoundError: No module named 'yolox'**

#!export PYTHONPATH="${PYTHONPATH}:/home/tpe-aa-02/AA/Jay/YOLOX"
!export PYTHONPATH="${PYTHONPATH}:/mnt/hdd1/AA/Jay/YOLOX"


## Appendix C: training config
1. 選擇GPU
2. 使用多張GPU卡

## Appendix D: 標注格式轉換
1. VOC to COCO
- https://blog.csdn.net/m_buddy/article/details/90348194
- http://10.109.6.14:9999/notebooks/x2coco/voc2coco.ipynb
2. YOLO 2 COCO
- http://10.109.6.14:9999/notebooks/x2coco/yolo2coco.ipynb

## Appendix E: 標注檔合併
- http://10.109.6.14:9999/notebooks/x2coco/combine_json.ipynb#Combine

# 1. Read Face & Person JSON
# 2. Combine Categories
# 3. Combine Annotation
# 4. Combin Images
# 5. Save Annotation

## Appendix E: Raspberry pi tips
- 安裝pytorch
- https://blog.csdn.net/phker/article/details/118190816
- https://torch.kmtea.eu/whl/stable-cn.html
- onnx export
- https://github.com/Megvii-BaseDetection/YOLOX/tree/main/demo/ONNXRuntime

python3 tools/export_onnx.py --output-name yolox_nano_dct.onnx -f exps/test/nano_dct.py -c YOLOX_outputs/nano_dct/best_ckpt.pth

- 轉ncnn
- https://github.com/Megvii-BaseDetection/YOLOX/tree/main/demo/ncnn/cpp
- https://zhuanlan.zhihu.com/p/391788686
- python執行ncnn inference
- 參考yolov5的作法: https://github.com/Tencent/ncnn/blob/master/python/ncnn/model_zoo/yolov5.py
- FPS Test
- Build ncnn yolox executable file
- https://blog.csdn.net/qq_39056987/article/details/119258569
- 編輯CMakeLists.txt

    vi /home/tpe-aa-01/AA/Jay/ncnn/examples/CMakeLists.txt
    加上 ncnn_add_example(yolox)

- 編輯yolox.cpp
- 更改model位置
- vulkan改成false

        vi /home/tpe-aa-01/AA/Jay/ncnn/examples/yolox.cpp

- 開始編譯

    cd /home/tpe-aa-01/AA/Jay/ncnn/build
    cmake ..
    make

- 產出: yolox執行檔

    ls /home/tpe-aa-01/AA/Jay/ncnn/build/examples/yolox

## Reference
1. pretrain weight: https://github.com/Megvii-BaseDetection/YOLOX
2. paper: https://arxiv.org/abs/2107.08430