tamnguyenvan / yolo-tf2

Yolov3 implementation in Tensorflow 2

YOLOv3 implementation in Tensorflow 2

An elegant Yolov3 implementation in Tensorflow 2.0.

What you need to know

This repo is heavily borrowed from awesome repo zzh8829. I just want to make it work on COCO 2017 dataset.

Installation

Clone the repo to your local

git clone https://github.com/tamnguyenvan/yolo-tf2

Install the requirements

pip install -r requirements.txt

Download yolov3 darknet weights and convert to tensorflow format

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

Usage

This is the time to enjoy. Let's detect some images!

python test.py --image_path /path/to/image --model_path ./checkpoints/yolov3.tf

Training

We also provide a pipeline for training the model on COCO 2017 dataset. Download COCO 2017 dataset, extract and put them into data/raw directory.

wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip

Create tfrecord files for training pipeline

python tools/create_tfrecords.py --data_dir ../data/raw --split train --output_file coco2017_train.tfrecord
python tools/create_tfrecords.py --data_dir ../data/raw --split val --output_file coco2017_val.tfrecord

The 2 tfrecord files should be generated in data/processed directory. If everything is done, let's train the model

python train.py \
	--train_file ./data/processed/coco2017_train.tfrecord \
	--val_file ./data/processed/coco2017_val.tfrecord \
	--batch_size 8 \
	--epochs 20 \
	--lr 0.001

References

This repo is heavily inspired by zzh8829. Don't forget to give him a star.

yolo-tf2's Issues

TypeError: transform_images() takes 1 positional argument but 2 were given

In test.py, line 35:

image = transform_images(image_raw, image_size)

However, the function transform_images defined in dataset.py takes only one positional argument:

def transform_images(image):
image /= 255.
return image