YAD2K: Yet Another Darknet 2 Keras

Welcome to YAD2K

You only look once, but you reimplement neural nets over and over again.

YAD2K is a 90% Keras/10% Tensorflow implementation of YOLO_v2.

Original paper: YOLO9000: Better, Faster, Stronger by Joseph Redmond and Ali Farhadi.

Requirements

Keras
Tensorflow
Numpy
Pillow (For rendering test results.)
Python 3

Installation

git clone https://github.com/allanzelener/yad2k.git
cd yad2k

# [Option 1] To replicate the conda environment:
conda env create -f environment.yml
source activate yad2k
# [Option 2] Install everything globaly.
pip install numpy

pip install tensorflow-gpu  # CPU-only: conda install -c conda-forge tensorflow
pip install keras # Possibly older release: conda install keras

Until Keras PR 5350 is merged:

pip install git+https://github.com/allanzelener/keras.git@func_defaults_to_tuple

Quick Start

Download Darknet model weights from the official YOLO website.
Convert the Darknet YOLO_v2 model to a Keras model.
Test the converted model on the small test set in images/.

wget http://pjreddie.com/media/files/yolo.weights
./yad2k.py cfg/yolo.cfg yolo.weights model_data/yolo.h5
./test_yolo.py model_data/yolo.h5  # output in images/out/

See ./yad2k.py --help and ./test_yolo.py --help for more options.

More Details

The YAD2K converter currently only supports YOLO_v2 style models, this include the following configurations: darknet19_448, tiny-yolo-voc, yolo-voc, and yolo.

yad2k.py -p will produce a plot of the generated Keras model. For example see yolo.png.

YAD2K assumes the Keras backend is Tensorflow. In particular for YOLO_v2 models with a passthrough layer, YAD2K uses tf.space_to_depth to implement the passthrough layer. The evaluation script also directly uses Tensorflow tensors and uses tf.non_max_suppression for the final output.

voc_conversion_scripts contains two scripts for converting the Pascal VOC image dataset with XML annotations to either HDF5 or TFRecords format for easier training with Keras or Tensorflow.

yad2k/models contains reference implementations of Darknet-19 and YOLO_v2.

Known Issues and TODOs

Error deserializing Lambda wrapping space_to_depth. Apply this PR to Keras.
- pip install git+https://github.com/allanzelener/keras.git@func_defaults_to_tuple
Add YOLO_v2 loss function.
Script to train YOLO_v2 reference model.
Support for additional Darknet layer types.
Tuck away the Tensorflow dependencies with Keras wrappers where possible.
YOLO_v2 model does not support fully convolutional mode.

Darknets of Yore

YAD2K stands on the shoulders of giants.

🔥 Darknet 🔥
Darknet.Keras - The original D2K for YOLO_v1.
Darkflow - Darknet directly to Tensorflow.
caffe-yolo - YOLO_v1 to Caffe.

K-Du / YAD2K