DockerCon EU 2017 project
To run the custom network written in C++ and OpenCL, open the project in Xcode and hit run. That easy. The project will then proceed to begin training via 'overfitting' on a single image to prove the loss function decreases over iterations, however this can easily be changed to train on numerous different images.
In order to run the tensorflow network, switch to the tf_network branch, and run python network_dist_train.py
. This will train the network using all the images stored in ./data
. (this folder needs to be filled first)
Upon running, the network will train, storing data in the folder ./train_dir
. (in some cases this folder will not be created automatically so just create a new folder if this happens)
Once the network is trained, store some images in a folder named ./out
, and then run python network_eval.py
.