nevers / space-dl

A convolutional neural network inspired by the YOLO algorithm to support spacecraft docking in space.

space-dl

A convolutional neural network inspired by the YOLO algorithm to support spacecraft docking in space.

The working principles and results are detailed and discussed in this blog post: [TODO]

Quickstart

Obtain the YOLOv3 Keras model

1. Download YOLOv3 model description and weights from the YOLO website.
2. Convert the Darknet YOLO model description to a Keras model using this script.

$ wget https://pjreddie.com/media/files/yolov3.weights
$ wget https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg
$ python convert.py yolov3.cfg yolov3.weights .model/yolov3.h5

Train and evaluate the model

1. Install python pip dependencies (consider using virtualenv for that).

$ virtualenv .venv
$ . .venv/bin/activate
$ pip install -r requirements

2. Simply run yolo3.py to start training and evaluation. It will automatically perform the following operations:

• download the dataset to ~/.datasets
• load the the pretrained model from yolov3.h5
• retrain the second half of the model whilst keeping the first half of the model parameters constant on your GPU.
• evaluate the model 2 times per epoch on your GPU
• write training and evalation metrics for tensorboard to .model/{train,eval}
• write positive and negative evaluation images, annotated with their predictions to .model/eval-img
• periodically save the model state to .model/model

$ ./yolo3.py

2. Run tensor board to monitor the training and evaluation process.

$ tensorboard --host=0.0.0.0 --logdir=.model

Point your browser to http://localhost:6006