YOLOv3, short for You Only Look Once version 3, is a state-of-the-art, real-time object detection algorithm that can detect multiple objects in an image or a video stream with remarkable speed and accuracy. Developed by Joseph Redmon and Ali Farhadi, YOLOv3 is a significant improvement over its predecessor, YOLOv2, with several innovative features such as a new backbone network, better object detection, and improved training techniques. Its architecture uses a single convolutional neural network (CNN) that simultaneously predicts bounding boxes and class probabilities for each object in the image, making it faster and more efficient than traditional object detection algorithms. YOLOv3 has a wide range of applications, including surveillance, autonomous vehicles, and robotics. In this blog, we will explore how to train YOLOv3 for custom object detection, a crucial task in many computer vision applications.
The notebook's GPUs must first be enabled:
- Select Notebook Settings under Edit.
- choose GPU using the Hardware Accelerator drop-down
Next, we'll check if Tensorflow can connect to the GPU: By executing the following code, you may quickly determine whether the GPU is enabled.
# Check if NVIDIA GPU is enabled
!nvidia-smi
Mounting the drive is the process of making an external storage device, such as a hard drive or cloud storage, accessible to the operating system. This allows us to store and load files from the mounted drive as if they were stored locally on the computer.
from google.colab import drive
drive.mount('/content/gdrive')
Let's make the required files and directories for training with custom objects.
- YOLOV3_Custom/images
- YOLOV3_Custom/custom.names
- YOLOV3_Custom/train.txt
- YOLOV3_Custom/test.txt
- YOLOV3_Custom/backup
- YOLOV3_Custom/detector.data
- YOLOV3_Custom/cfg
Changing directory to drive Directory
# changing directory to the google drive
import os
drive_path = os.path.join(os.getcwd(), "gdrive/MyDrive")
%cd {drive_path}
Assigning path of the home directory to HOME for easy file handling
HOME = os.path.join(drive_path, "YOLOV3_Custom")
HOME
Creating Home directory to store all the files
os.mkdir(f"{HOME}")
Changing current directory to HOME
%cd {HOME}
We will train our model to recognise pistols in this project, thus we must gather the images and its annotaions and save them in the We will train our model to recognise pistols in this project, thus we must gather the images and its annotaions and save them in the YOLOV3_Custom/images directory.
If you download the dataset from the 1st link, then no need to create image directory, just download the zip file into the YOLOV3_Custom directory and unzip it.
Your Google Drive file location that you wish to unzip is the path in the cell below.
In our case images and text files should be saved in YOLOV3_Custom/images directory. For e.g. image1.jpg should have a text file image1.txt.
%cd {HOME}
!unzip "zip/images.zip"
Labels of our objects should be saved in YOLOV3_Custom/custom.names file, each line in the file corresponds to an object. In our case since we have only one object class, the file should contain the following.
custom.names
Pistol
The annotated photos can then be randomly split into train and test sets in a 80:20 ratio.
YOLOV3_Custom/train.txt Inside the train.txt paths of the 80% images should be listed.
YOLOV3_Custom/test.txt Inside the test.txt paths of the 20% images should be listed. train.txt
images/armas (1000).jpg
images/armas (1001).jpg
images/armas (1002).jpg
images/armas (1003).jpg
images/armas (1004).jpg
images/armas (1005).jpg
images/armas (1006).jpg
images/armas (1007).jpg
To divide all image files into 2 parts. 90% for train and 10% for test, Upload the process.py in YOLOV3_Custom directory
This process.py script creates the files train.txt & test.txt where the train.txt file has paths to 90% of the images and test.txt has paths to 10% of the images.
You can download the process.py script from my GitHub.
Open process.py specify the paths and then run it.
%cd {HOME }
# run process.py ( this creates the train.txt and test.txt files in our darknet/data folder )
!python process.py
# list the contents of data folder to check if the train.txt and test.txt files have been created
!ls
Creating backup directory for storing weights of the trained model.
%cd {HOME}
os.mkdir("backup")
detector.data
classes=1
train=../train.txt
valid=../test.txt
names=../custom.names
backup=../backup
The detector.data file is a configuration file for training an object detection model. Here is what each line of the file means:
- classes=1: This line indicates the number of classes or objects that the model will detect. In this case, the model will detect only one class.
- train=../train.txt: This line specifies the path to the text file containing the paths to the training images.
- valid=../test.txt: This line specifies the path to the text file containing the paths to the validation images.
- names=../custom.names: This line specifies the path to the file containing the names of the classes that the model will detect.
- backup=../backup: This line specifies the path where the weights of the trained model will be saved during the training process.
Darknet, an open source neural network framework, will be used to train the detector. creating a local copy of the Darknet repository on your computer, which you can then use to build and train neural network models for object detection.
%cd {HOME}
!git clone https://github.com/AlexeyAB/darknet# Change current working directory to Darknet
%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
- The first command replaces the string OPENCV=0 with OPENCV=1 in the Makefile. This is done to enable OpenCV support in Darknet, which is necessary for some image-related tasks.
- The second command replaces the string GPU=0 with GPU=1 in the Makefile. This is done to enable GPU acceleration in Darknet, which can greatly speed up training and inference.
- The third command replaces the string CUDNN=0 with CUDNN=1 in the Makefile. This is done to enable cuDNN support in Darknet, which is an NVIDIA library that provides faster implementations of neural network operations.
- The fourth command replaces the string CUDNN_HALF=0 with CUDNN_HALF=1 in the Makefile. This is done to enable mixed-precision training in Darknet, which can further speed up training and reduce memory usage.
The !make command is a Linux command-line instruction that invokes the make utility to compile and build the Darknet codebase based on the configurations specified in the Makefile. This command reads the Makefile in the current directory and compiles the source code by executing various build commands specified in the Makefile. After the compilation process is complete, the make utility generates an executable binary file that can be used to run various Darknet commands and utilities.
# compiling the model
!make
We can choose the YOLOv3 configuration file based on the performance that is necessary. Yolov3.cfg will be used in this example. The file from darknet/cfg/yolov3.cfg can be copied to YOLOV3_Custom/cfg/yolov3-custom.cfg.
_
The neural network weights are iteratively adjusted as the images are trained. We may use large training sets, making it resource-intensive to update the weights for the full training set in a single cycle. The batch parameter is specified to utilise a small set of images to iteratively update the weights. It is initially set at 64.
_
The parameter max batches determines the maximum number of iterations for which our network should be trained. You can calculate its value with the formula max_batches = classes*2000, i.e. 2000 for a single class, and modify the line steps to 80% and 90% of max batches, which is 1600,1800.
when the model reaches to 2000 epochs during training for 1 class, it will stop the training, because we defined that in our max_batches parameter, if you want to train your model beyound from that, increase the value of max_batches, and also increase its steps, like when you change the value to max batches=6000, modify the line steps to 80% and 90% of max batches, i.e. steps=4800,5400.
--
The classes and filters parameters of the [yolo] and [convolutional] layers immediately preceding the [yolo] layers must be updated.
Since there is just one class in this project (Pistol), we will update the class parameter in the [yolo] layers to 1 at lines 610, 696, and 783.
The filters parameter will also need to be updated based on the classes count: filters=(classes + 5) * 3. We should set filters=18 for a single class at line numbers: 603, 689, 776.
YOLOV3_Custom/cfg/yolov3-custom.cfg contains all configuration changes.
To train our object detector, we can use the pre-trained weights that have already been trained on a large data sets. The pre-trained weights are available here, and they can be downloaded to the root directory.
%cd {HOME}/pt-weights
# Download weights darknet model 53
!wget https://pjreddie.com/media/files/darknet53.conv.74
# changing current drive to the darknet
%cd {HOME}/darknet
To start training a custom object detector in Darknet, we need to have all the required files and annotated images. Once we have them, we can begin the training process and continue it until the loss reaches a predefined level. During training, the weights of the detector are saved periodically to disk. Initially, they are saved once every 100 iterations until 1,000 iterations, after which they are saved once every 10,000 iterations. After training is complete, we can use the generated weights for object detection.
!./darknet detector train {HOME}/detector.data {HOME}/cfg/yolov3-custom.cfg {HOME}/pt-weights/darknet53.conv.74 -dont_show
If your model training is interrupted for any reason, such as a notebook timeout, a notebook crash, network issues, or other reasons, you can continue training from where you left off. This is done by passing the previously trained weights to the training process, which allows the model to pick up where it left off and continue improving.
!./darknet detector train {HOME}/detector.data {HOME}/cfg/yolov3-custom.cfg {HOME}/backup/yolov3-custom_4000.weights -dont_show
As we have trained our weights up to 2000 epochs, we will calculate the mean average precision of our trained weights.
!./darknet detector map {HOME}/detector.data {HOME}/cfg/yolov3-custom.cfg {HOME}/backup/yolov3-custom_4000.weights -dont_show
Make changes to your custom config file
- change line batch to batch=1
- change line subdivisions to subdivisions=1
You can do it either manually or by simply running the code below
#set your custom cfg to test mode
%cd {HOME}/cfg
!sed -i 's/batch=64/batch=1/' yolov3-custom.cfg
!sed -i 's/subdivisions=16/subdivisions=1/' yolov3-custom.cfg
# define helper function imShow
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()
%cd {HOME}/darknet
# run your custom detector with this command (upload an image to your google drive to test, the thresh flag sets the minimum accuracy required for object detection)
!./darknet detector test {HOME}/detector.data {HOME}/cfg/yolov3-custom.cfg {HOME}/backup/yolov3-custom_2000.weights {HOME}/pistol_Image.jpg -thresh 0.3
%cd {HOME}/darknet
imShow('predictions.jpg')
For inference use the python file which is inside the inference directory.