Refine online augmentation

[scottclowe]

The set of augmentations I rustled up from "looking at what was generated by eye" keeps doing worse than online 8 aug.

We need to work if one of the augmentations is hindering the learning.

Generally, we need to find the best set of augmentations to use with this dataset.

[scottclowe]

I'm running models which I have tried to optimise to run faster and I'm comparing

• 8 aug (4 rotations, + flip)
• uniform rotations + flip
• uniform rotations + flip + shear N(0, 5deg)
• uniform rotations + flip + scale N(100%, 10%)
• uniform rotations + flip + translation N(0, 7.5%) [both x and y]

Shorthand: N(mu,sigma) = normal distribution

[scottclowe]

Translation seemed to be a bit of a hindrance, whereas shear and scale were not. In all cases, augmentations prevent overfitting - keeping train_nll at the same level as valid_nll.

[scottclowe]

Comparing transform_order, it seems than the default of 0.5 is slightly better than using linear interpolation (1.0) after the same number of epochs. However, training on the test 10% of data it is ~5% slower as well.

The difference in duration on the train 80% of data will be much larger, so maybe we should train with transform_order:1 for the speed benefits now we are short on time... We could use transform_order:1 for the bulk of the training and then transform_order:0.5 for fine tuning at the end of training?

[scottclowe]

The final difference was a logloss validation of about 1.6 vs 1.9 .