YOLOv3: Inplement in Tensorflow 2

This repo is the implementation of YOLOv3 with Tensorflow 2. It refers to many repos as mentioned in Acknowledgments.

Original paper: YOLOv3: An Incremental Improvemen by Joseph Redmon and Ali Farhadi.

Darknet: https://github.com/pjreddie/darknet

Step-by-Step

yolov3_step_by_step.ipynb

1. Jupyter Notebook

   $ jupyter notebook

2. Colab

   

Prepare the Dataset

1. Please download the VOC2012 dataset and put it into data folder.

   $ mkdir data
   $ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar -O ./data/VOCtrainval_11-May-2012.tar
   $ tar xvf ./data/VOCtrainval_11-May-2012.tar --directory ./data
   

2. Split dataset and transfer to tfrecord.

   # train
   $ python3 voc2012.py \
       --data_dir ./data/VOCdevkit/VOC2012/ \
       --split train \
       --output_dir ./data
   
   # val
   $ python3 voc2012.py \
       --data_dir ./data/VOCdevkit/VOC2012/ \
       --split val \
       --output_dir ./data

3. Validate the dataset

   $ python3 visualize.py

Training

Training from scratch

$ python3 train.py --transfer=Fasle

Transfer Learning

1. Download pre-trained Darknet weights

   $ wget https://pjreddie.com/media/files/yolov3.weights -O model_data/yolov3.weights

   Testing pre-trained Darknet weights

   

2. Convert Darknet weights to Tensorflow weights

   $ mkdir -p checkpoints
   $ python3 convert.py

3. Validate pre-trained weight

   $ python3 detect.py

4. Training

   $ python3 train.py \
       --size 416 \
       --epochs 10 \
       --num_classes 20 \
       --batch_size 16 \
       --train_dataset ./data/voc2012_train.tfrecord \
       --val_dataset ./data/voc_2012_val.tfrecord \
       --transfer=True \
       --pretrained_weights ./checkpoints/yolov3.tf \
       --weights_num_classes 80

Inference

1. Inference with pre-trained Darknet weights

   $ python3 detect.py

2. Inference with the model that you just trained

   $ python3 detect.py \
       --classes ./model_data/voc2012_classes.txt \
       --num_classes 20 \
       --weights=./checkpoints/yolov3_train_10.tf \
       --image=./data/street.jpg

   Testing the model that we just trained through 10 epochs

   

3. Other command option please using --help to see, as below:

   $ python3 detect.py --help

References

• pjreddie/darknet

  • official yolov3 implementation

• YAD2K

  • models

• qqwweee/keras-yolo3

  • models
  • loss functions

• yolov3-tf2

  • models
  • loss functions