ENAS integration

[ahundt]

Start your first task with my fork of the ENAS repository:
https://github.com/ahundt/enas/tree/good_fixes
Modify the cifar10 code to also support running fashion-mnist
Make sure it is possible to provide a tensorflow tensor with input values
This part may take some time to understand. We can discuss how it works when you have time.
Make sure to support variable dimension inputs, since we plan to run on robot datasets next which can have 3-4 channels (red green blue depth).

[ahundt]

Almost there!

we might want to ask the following questions: 1. can cells be created separately fairly easily with the current code? 2. Can fashion mnist be trained well with the settings we’re using for grasping due to memory limitations (ex: 3 cells instead of 5)?

ahundt [7:15 PM]
regarding 1: We may want to consider supplying each image to a couple of cells each, concatenating that with the pose information, and passing the result of that to the next cell
ahundt [7:16 PM]
4: the convolutions may only have 20 filters, since we are inputting 128x128x15 values, should we consider increasing the number of filters for the current configuration?
ahundt [7:16 PM]
5: can any of the hyperopt models be trained from scratch (no imagenet weights)?
ahundt [7:18 PM]
6: could the principles of the paper yotam linked be causing us regression accuracy problems https://eng.uber.com/coordconv/