Lots of new updates, testing can be done now with results saved to Google Cloud Storage.
Class level hard example mining, triplet formation and class rectification loss are working. I have debugged the distributed_job to the point where it will run for an epoch but I couldn't figure out how to easily save model weights with the net in distributed mode. This is a todo item.
Updated the train script to run for 25 epochs and found the optimal starting learning rate in the Kaggle kernel. A lot of documentation and refactoring to do still. Need to give paper authors credit.
Lots of new updates, testing can be done now with results saved to Google Cloud Storage.
Class level hard example mining, triplet formation and class rectification loss are working. I have debugged the distributed_job to the point where it will run for an epoch but I couldn't figure out how to easily save model weights with the net in distributed mode. This is a todo item.
Updated the train script to run for 25 epochs and found the optimal starting learning rate in the Kaggle kernel. A lot of documentation and refactoring to do still. Need to give paper authors credit.