Closed gngdb closed 9 years ago
Also running one with just more rotations, up to 8 rotations, should use 12% of the RAM.
Completed this and it did improve our score. Had to hack in alternative dataset loading into test.py
to accomodate making predictions using it though.
Worth checking to see if we continue to get returns on augmentation in RAM. Running a new model with crops as well as rotations and flips and loading into RAM. That makes 40 augmentations in total. Should be about 56% of the RAM doing this. Running now to see what happens.