Experiment: Visual Data Augmentation

[rowandempster]

Two settings:

• No augmentation to images we load
• Standard image augmentation pipeline (https://pytorch.org/vision/main/transforms.html# https://pytorch.org/vision/main/auto_examples/plot_transforms.html#sphx-glr-auto-examples-plot-transforms-py)

Note that some augmentations will also affect the RoIs.

Code to put transforms into:

• Perception-Research/src/PGP/datasets/visual_data_helper.py

  Line 22 in f256bfe

  def _compose_transform(self, transform: Dict) -> transforms.Compose:

• Perception-Research/src/PGP/configs/pgp_gatx2_lvm_traversal_visual.yml

  Line 7 in f256bfe

  common_transform: &common_transform

Remember that some transforms only should apply at training time

Result of this ticket should be tensorboard logs of 100epochs under the two settings, with none of the other code changed beside the setting