Add PAE ranking and image directory data loader

[bdubayah]

• Add new file for standard PAE outlier detection
• Add new file for convolutional PAE outlier detection
  • Create tensorflow datasets from lists of images
• Add data loader that loads a directory of images (for convolutional PAE)
  • Update base outlier detection class to not try to cast the directory listings being passed through to float32
• Add FMNIST and MNIST data as sample data with an experiment configured for running the convolutional PAE

[bdubayah]

Hi @hannah-rae and @stevenlujpl , the PAE ranking modules are now finalized and ready for review – I edited the top comment to include all the major changes. Let me know if there's any questions or changes needed!

[stevenlujpl]

@bdubayah Thanks for making the changes, and they all look great to me.

I still have some questions:

1. If I run the PAE script, will the training process take place on CPU or GPU?
2. This problem may already be resolved in tensorflow 2. When I used tensorflow 1 a long time ago, tensorflow 1 would hold on to all the available GPU memories (even it doesn't really need) in a machine. This behavior is annoying (especially when using it on a shared machine). If this problem isn't resolved in tensorflow 2, can you please configure tensorflow to use only the memory it needs?

[stevenlujpl]

This is how I limited the GPU memory usage in tensorflow 1.x:

import tensorflow as tf

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config, ...)

[stevenlujpl]

Another option is to set the visible GPUs to be a single GPU.

[bdubayah]

Good catch, I didn't know that tensorflow did that - it will run on GPUs if available. Maybe we could by default set it to one gpu with growth on and then add config options to allow for multiple gpus and enabling/disabling growth? Or an option to entirely disable GPUs?