Darknet yolov4-tiny (feat. Tensorflow Lite)

The $root path is assumpted as current repository path.
This repository includes a setup for trianing person detector, but you can apply this method for training the other dataset.

Compile Darknet

1. Clone 'darknet' git repository to the current $root.

~$ git clone https://github.com/AlexeyAB/darknet 

2. Change makefile to have GPU and OPENCV enabled # also set CUDNN, CUDNN_HALF and LIBSO to 1

~$ cd ./darknet/

~$ sed -i 's/OPENCV=0/OPENCV=1/' Makefile
~$ sed -i 's/GPU=0/GPU=1/' Makefile
~$ sed -i 's/CUDNN=0/CUDNN=1/' Makefile
~$ sed -i 's/CUDNN_HALF=0/CUDNN_HALF=1/' Makefile
~$ sed -i 's/LIBSO=0/LIBSO=1/' Makefile

3. Build darknet

~$ sudo apt install libopencv-dev   # (ref) https://github.com/pjreddie/darknet/issues/2280

~$ make 

4. Clean the data and cfg folders first except the labels folder in data which is required

~$ cd ./darknet/data 
~$ find -maxdepth 1 -type f -exec rm -rf {} \;

~$ cd .. 
~$ rm -rf cfg/ 
~$ mkdir cfg 

5. Unzip the datasets and their contents so that they are now in /darknet/data/ folder

• you can preprare your own customized dataset
• if you want to know how to prepare custom dataset, refer to this article.

~$ unzip ./obj.zip -d ./darknet/data 

6. Copy the custom cfg file from the drive to the darknet/cfg folder

~$ cp ./yolov4-tiny-custom.cfg ./darknet/cfg 

7. Copy the obj.names and obj.data files so that they are now in /darknet/data/ folder

~$ cp ./yolov4-tiny/obj.names ./darknet/cdata
~$ cp ./yolov4-tiny/obj.data  ./darknet/data

8. Copy the process.py file from the drive to the darknet directory

~$ cp ./process.py ./darknet/

9. Run the process.py python script to create the train.txt & test.txt files inside the data folder.

~$ cd ./darknet/ 

~$ python process.py      # this creates the train.txt and test.txt files in our darknet/data folder
~$ ls ./darknet/data/     # list the contents of data folder to check if the train.txt and test.txt files have been created 

10. Download the pre-trained yolov4-tiny weights.

~$ cd ./darknet/ 
~$ wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29  # Download the yolov4-tiny pre-trained weights file

---

Training

train your custom detector!

~$ cd ./darknet/ 
~$ mkdir ./darknet/data/training     # train checkpoint will be saved here 

~$ ./darknet detector train ./data/obj.data cfg/yolov4-tiny-custom.cfg yolov4-tiny.conv.29 -dont_show -map   # run training 

(option) to restart training your custom detector where you left off(using the weights that were saved last) (히스토리에서 가져올 때)

~$ ./darknet detector train ./data/obj.data cfg/yolov4-tiny-custom.cfg ./data/training/yolov4-tiny-custom_last.weights -dont_show -map  # re-train from the checkpoint 

check performance

You can check the mAP for all the saved weights to see which gives the best results ( xxxx here is the saved weight number like 4000, 5000 or 6000 snd so on )

~$ cd ./darknet 
~$ ./darknet detector map ./data/obj.data ./cfg/yolov4-tiny-custom.cfg ./data/training/yolov4-tiny-custom_best.weights -points 0

---

Convert Darknet to TensorFlow

1. Clone 'Tensorflow light for yolov4'

~$ git clone https://github.com/hunglc007/tensorflow-yolov4-tflite.git
~$ cd ./tensorflow-yolov4-tflite
~$ mkdir checkpoints
~$ pip install -r requirements.txt 

2. Convert the weights to TensorFlow's .pb representation

~$ cp ./darknet/data/obj.names ./tensorflow-yolov4-tflite/data/classes/ 

3. Change the labels from the default COCO to your own custom ones.

~$ sed -i "s/coco.names/obj.names/g" ./tensorflow-yolov4-tflite/core/config.py

4. Convert darknet weights to tensorflow

Convert to both a regular TensorFlow SavedModel and to TensorFlow Lite.
For TensorFlow Lite, we'll convert to a different TensorFlow SavedModel beforehand.

~$ cd ./tensorflow-yolov4-tflite

~$ python save_model.py \
  --weights ../darknet/data/training/yolov4-tiny-custom_best.weights \
  --output ./checkpoints/yolov4-tiny-416 \
  --input_size 416 \
  --model yolov4 \
  --tiny \

Run demo tensorflow

~$ pip install -U opencv-python==4.1.2.30
~$ python detect.py --weights ./checkpoints/yolov4-tiny-416 --size 416 --model yolov4 --image ./data/girl.png --tiny True --output 'result.png' --score 0.7   # check the results 

Convert the Darknet model weights to TensorFlow Lite

~$ cd ./tensorflow-yolov4-tflite

~$ python save_model.py \
  --weights ../darknet/data/training/yolov4-tiny-custom_best.weights \
  --output ./checkpoints/yolov4-tiny-416-tflite\
  --input_size 416 \
  --model yolov4 \
  --tiny \
  --framework tflite

# # SavedModel to convert to TFLite
~$ python convert_tflite.py --weights ./checkpoints/yolov4-tiny-416-tflite --output ./checkpoints/yolov4-tiny-416.tflite

# Run demo tensorflow 
~$ python detect.py --weights ./checkpoints/yolov4-tiny-416.tflite --size 416 --model yolov4 --image ./data/girl.png --framework tflite --tiny --score 0.2

Convert the TensorFlow weights to TensorFlow Lite

~$ cd ./tensorflow-yolov4-tflite 

~$ python convert_tflite.py --weights ./checkpoints/yolov4-tiny-416-tflite  --output ./checkpoints/yolov4-tiny-416.tflite

---

Reference

[1] AVA-Dataset-Processing-for-Person-Detection / for training person detection dataset
[2] yolov4-tiny-tflite-for-person-detection / an example of person detector trained by Darknet
[3] TRAIN A CUSTOM YOLOv4-tiny OBJECT DETECTOR USING GOOGLE COLAB /
[4] tensorflow-yolov4-tflite / to convert darknet to tensorflow