I see you have clearly stated that pba has been tested with tensorflow 1.10.0 and 1.11.0 and so is the case with ray which has been tested with 0.7.0 version. It'd be great, it can be added to the requirements.txt file as, for example, tensorflow-gpu=1.11.0. Also, does it make sense to add tensorflow in the requirements file as well for cpu based execution?

In pil_unwrap of augmentation_transforms.py, the code takes in only square images. Is there a possibility to handle images of unequal width and height?