Tutorial: Yolov3 in Keras

Reference

Github: https://github.com/qqwweee/keras-yolo3

Setup

Create Virtual Environment (Conda)

Lets create a virtual environment for YOLOv3.

The requirements are

• python=3.7.10

• cudatoolkit=10.0

• cudnn-7.6.5-cuda10.0_0

• tensorflow-gpu=1.15.0

• keras=2.3.1

• pillow=8.2.0

• matplotlib=3.3.4

• opencv=3.4.2

If you have problems when installing opencv packages, use the following commands pip install opencv-python

conda create -n tf115 python=3.7
conda activate tf115

Install the following:

conda install cudatoolkit=10.0
conda install cudnn
conda install tensorflow-gpu=1.15.0
conda install keras=2.3
conda install pillow
conda install matplotlib
conda install opencv

Clone Git

After the installation, activate the virtual environment. We will clone the reference repository to download Yolov3 codes.

Method 1: From conda prompt (in virtual env)

git https://github.com/qqwweee/keras-yolo3.git

Method 2:

Download zip file from the github and unzip.

Download the trained weight file

After the download, place the weight model file in the same directory of Yolov3.

• YOLOv3 weights https://pjreddie.com/media/files/yolov3.weights

You can also download it from the conda Prompt as

wget https://pjreddie.com/media/files/yolov3.weights``

• YOLOv3-tiny weights

https://pjreddie.com/media/files/yolov3-tiny.weights

Open V.S Code

>> code .

You can also run the below codes in the Conda Promt

In VS code, select the virtual environment: F1--> Python Interpreter --> Select Environ.

Convert Darknet YOLOv3 to Keras model

In the terminal of VS code or in Conda Prompt, type:

>> python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

Run Yolov3 Detection

Copy the test video file in the same directory (Yolov3 directory)

If the video file name is 'test_Video.avi'

>> python yolo_video.py --model .\model_data\yolo.h5 --input .\test_Video.avi

Run Yolov3-Tiny Detection

After downloading yolov3-tiny.weights, Convert it to Keras model and save it as 'yolo-tiny.h5'

 >> python convert.py yolov3-tiny.cfg yolov3-tiny.weights model_data/yolo-tiny.h5

Run Yolo-tiny with the Test video

>> python yolo_video.py --model .\model_data\yolo-tiny.h5 --input .\test_Video.avi

Usage

Use --help to see usage of yolo_video.py:

usage: yolo_video.py [-h] [--model MODEL] [--anchors ANCHORS]
                     [--classes CLASSES] [--gpu_num GPU_NUM] [--image]
                     [--input] [--output]

positional arguments:
  --input        Video input path
  --output       Video output path

optional arguments:
  -h, --help         show this help message and exit
  --model MODEL      path to model weight file, default model_data/yolo.h5
  --anchors ANCHORS  path to anchor definitions, default
                     model_data/yolo_anchors.txt
  --classes CLASSES  path to class definitions, default
                     model_data/coco_classes.txt
  --gpu_num GPU_NUM  Number of GPU to use, default 1
  --image            Image detection mode, will ignore all positional arguments

How to train a dataset

Prepare the dataset

For this tutorial, we will use KITTI dataset

• Image file: Download Kitti Dataset
• Label file: Download Kitti Dataset Label

  • Object Detection annotation Convert to Yolo Darknet Format: Click here
• Class file:

  • Copy the 'kitti_classes.txt' in the folder of `\model_data` folder

Modify train.py

Open 'train.py' file in VS Code\

Go to LIne 16 : def main():. Change the ''annotation' and 'classes-path' to your setting.

def _main(): 
annotation_path = 'train.txt' log_dir = 'logs/000/'
#classes_path = 'model_data/voc_classes.txt'
classes_path = 'model_data/kitti_classes.txt'
anchors_path = 'model_data/yolo_anchors.txt'

Go to LIne 32: Change the name of the pre-trained weight file.

• We will use COCO trained weight file as we used above(yolo.h5). Create a copy and name it asyolo_weights.h5

    if is_tiny_version:
        model = create_tiny_model(input_shape, anchors, num_classes,
            freeze_body=2, weights_path='model_data/yolo_tiny_weights.h5')
    else:
        model = create_model(input_shape, anchors, num_classes,
            freeze_body=2, weights_path='model_data/yolo_weights.h5') # make sure you know what you freeze

Run Train

Start training by running the following in the terminal

>>python train.py

Evaluate

Use your trained weights or checkpoint weights with command line option --model model_file when using yolo_video.py Remember to modify the class path or anchor path.

TroubleShooting

Problem 1

Error message of

_, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])

Solution

Modify model.py (line 394)

_, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])

should be changed to

_, ignore_mask = tf.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])