Average Time Elapsed for process on each frame: 71.6ms

YoloV3 Implementation in Tensorflow 2.x

See on YouTube: 4K YoloV3 Object detection using Tensorflow 2.1

This repo provide a clean implementation of YoloV3 in Tensorflow 2.x using all the best practices.

Key Features

• TensorFlow 2.1
• yolov3 with pre-trained Weights
• yolov3-tiny with pre-trained Weights
• Inference example
• Transfer learning example
• Eager mode training with tf.GradientTape
• Graph mode training with model.fit
• Functional model with tf.keras.layers
• Input pipeline using tf.data
• Tensorflow Serving
• Vectorized transformations
• GPU accelerated
• Fully integrated with absl-py from abseil.io
• Clean implementation
• Following the best practices
• MIT License

Usage

Installation

Pip

# without GPU
pip install -r requirements.txt
#with GPU
pip install -r requirements-gpu.txt

Nvidia Driver (For GPU)

# Ubuntu 18.04
sudo apt-add-repository -r ppa:graphics-drivers/ppa
sudo apt install nvidia-driver-440
# Windows/Other
https://www.nvidia.com/Download/index.aspx

Convert pre-trainned Darknet weights

# yolov3
wget https://pjreddie.com/media/files/yolov3.weights -O data/yolov3.weights
python convert.py --weights ./data/yolov3.weights --output ./checkpoints/yolov3.tf

Detection

# WebCam
python object_detected.py --video 0

# Video file
python object_detected.py --video path_to_file.mp4

# Vidoe file with output
python object_detected.py --video path_to_file.mp4 --output ./output.avi

Training

soon

Implementation Details

Loading pre-trained Darknet weights

very hard with pure functional API because the layer ordering is different in tf.keras and darknet. The clean solution here is creating sub-models in keras. Keras is not able to save nested model in h5 format properly, TF Checkpoint is recommended since its offically supported by TensorFlow.

Command Line Args Reference

convert.py:
  --output: path to output
    (default: './checkpoints/yolov3.tf')
  --weights: path to weights file
    (default: './data/yolov3.weights')
  --num_classes: number of classes in the model
    (default: '80')
    (an integer)


detect_video.py:
  --classes: path to classes file
    (default: './data/coco.names')
  --video: path to input video (use 0 for cam)
    (default: './data/video.mp4')
  --output: path to output video (remember to set right codec for given format. e.g. XVID for .avi)
    (default: None)
  --output_format: codec used in VideoWriter when saving video to file
    (default: 'XVID')
  --weights: path to weights file
    (default: './checkpoints/yolov3.tf')
  --num_classes: number of classes in the model
    (default: '80')
    (an integer)

Change Log

May 25, 2020

• Updated to Tensorflow to v2.1.0 Release

References

It is pretty much impossible to implement this from the yolov3 paper alone.

• https://github.com/pjreddie/darknet
  • official yolov3 implementation
• https://github.com/AlexeyAB
  • explinations of parameters
• https://github.com/qqwweee/keras-yolo3
  • models
  • loss functions
• https://github.com/YunYang1994/tensorflow-yolov3
  • data transformations
  • loss functions
• https://github.com/ayooshkathuria/pytorch-yolo-v3
  • models
• https://github.com/broadinstitute/keras-resnet
  • batch normalization fix
• https://github.com/zzh8829/yolov3-tf2
  • tiny