Online augmentation support for check_test and test

[gngdb]

Currently, all the models we've been training have, for some reason, been using the 4D representation (channels, width,height,examples) for minibatches, in the YAML file. When writing the online augmentation I "fixed" this so that it would be the other way because it's more memory efficient : (examples,width,height,channels).

Unfortunately, the check_test.pu and test.py scripts aren't very smart and can't deal with the change. We need to have a check to see which way it is (don't want to break old models) and then run the forward pass with or without the transpose. The model file that's loaded should have a representation stored internally we can just pull out and check.

Also, we need to decide how many augmentations to average predictions over; or to have an option to not apply augmentations. So there are three changes:

• [x] add support for forward prop in transposed minibatches
• [x] add an option to indicate how many augmented runs through the test data to do
• [ ] add an option to run without any augmentation

[gngdb]

Oh yeah, and of course they both assume you're wanting to use just the traditional DensePNGDataset. So also:

• [x] Change dataset loaded based on model YAML (maybe just load the YAML as train does and take the dataset from there)

[gngdb]

Now dataset in check_test_score is loaded from the YAML with some clever parsing, should support both datasets. Tested with dense and it works fine.

[gngdb]

The above features are now implemented for check_test_score.py but not test.py, but it should be easy to copy most of the code across.

[gngdb]

There was a bug in check_test_score.py for a while there. Now fixed, but can no longer reproduce the score of 0.75 seen during model training.

[gngdb]

test.py now supports generating models that use ListDataset (online augmented) and models that use ParallelDataset. Both dataset files had to be updated for this to work.