planetfederal / thomas

thomas

About

Thomas is a training engine that runs on coal

Being cute aside, thomas is an object detection and segmentation model that we extend to detect objects in imagery - using Tensorflow, Keras amd Coco. While we train the models with new data, the majority of the hard work was done here. The base mrcnn model was copied from that repo and slightly modified to fix a bug or two. The code in the mrcnn directory should not be touched unless you really know what you are doing.

Our model code lives in the objects directory.

Building

The dev environment for this can be a little trick but works on my machine (osx) as well as on Amazon's deep learning AMI

To get up and running:

• Setup a python 3.6 virtual env and source it
• pip install -r requirements
• export PYTHONPATH=$PATHONPATH:.
• Get the precomputed model weights from S3, either s3://tomsorflow/coco/mask_rcnn_coco.h5 or s3://tomsorflow/buildings/mask_rcnn_buildings_1.h5

The setup_env.sh should do the trick most of the time

Training

Analyzing