Skip to main content

Train a Custom Object Detector

ยท 3 min read
Amir Afshari

download

Darknet Configurations

This documentation is for Google Colab. If you want to know how to compile darknet on your linux local machine (Ubuntu 20.04), please read this documentation.

# clone repo
#!git clone https://github.com/AlexeyAB/darknet
!git clone https://github.com/amirafshari/LPD-YOLOv4

GPU

# change makefile to have GPU and OPENCV enabled
%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

# verify CUDA
!/usr/local/cuda/bin/nvcc --version

# make darknet
!make

Weights

# pre-trained weights on MS COCO dataset
!wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

# pre-trained weights for the convolutional layers
!wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.conv.137

Generate train.txt and test.txt

These files are not in the official repo, but you can find them in my repository.

!python generate_train.py
!python generate_test.py

Configurations

We need to change/create these files (I configured them for our object (which is license plate), and put them in this repository):

  • data/obj.names
  • data/obj.data
  • cfg/yolov4-custom.cgf
  • cfg/yolov4-obj.cfg

Training

Configurations

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

  • 1 Epoch = images_in_train_txt / batch = 2000 / 32 = 62.5

Train

# Access Denied Error
!chmod +x ./darknet

# set custom cfg to train mode 
%cd cfg
!sed -i 's/batch=1/batch=64/' yolov4-obj.cfg
!sed -i 's/subdivisions=1/subdivisions=16/' yolov4-obj.cfg
%cd ..

!./darknet detector train ./data/obj.data ./cfg/yolov4-obj.cfg yolov4.conv.137 -dont_show -map

Restart

In case of intruption, we can restart training from our last weight.
(every 100 iterations our weights are saved to backup folder in yolov4-obj_last.weights) (~every 30 minutes)
(every 1000 iterations our weight are saved to backup folder in yolo-obj_xxxx.weights)

!./darknet detector train ./data/obj.data ./cfg/yolov4-obj.cfg ./backup/yolov4-obj_last.weights -dont_show -map

Sanity Check

Setup

# set custom cfg to test mode 
%cd cfg
!sed -i 's/batch=64/batch=1/' yolov4-obj.cfg
!sed -i 's/subdivisions=16/subdivisions=1/' yolov4-obj.cfg
%cd ..

def imShow(path):
  import cv2
  import matplotlib.pyplot as plt
  %matplotlib inline

  image = cv2.imread(path)
  height, width = image.shape[:2]
  resized_image = cv2.resize(image,(3*width, 3*height), interpolation = cv2.INTER_CUBIC)

  fig = plt.gcf()
  fig.set_size_inches(18, 10)
  plt.axis("off")
  plt.imshow(cv2.cvtColor(resized_image, cv2.COLOR_BGR2RGB))
  plt.show()

COCO Dataset

!./darknet detector test cfg/coco.data cfg/yolov4.cfg yolov4.weights data/person.jpg

imShow('./predictions.jpg')

download

Custom Dataset

!./darknet detector test ./data/obj.data ./cfg/yolov4-obj.cfg ./backup/yolov4-obj_last.weights ../Cars354.png -thresh 0.3

imShow('./predictions.jpg')

result-4

To process a list of images data/train.txt and save results of detection to result.json file use

!./darknet detector test data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -ext_output -dont_show -out result.json < data/test.txt

Metrics

Use -map flag while training for charts
mAP-chart (red-line) and Loss-chart (blue-line) will be saved in root directory.
mAP will be calculated for each 4 Epochs ~ 240 batches

!./darknet detector map data/obj.data cfg/yolov4-obj.cfg backup/custom.weights