Population Based Augmentation

Train a population of randomly initialized child classifiers and periodically compare their validation performance. In the published truncation-selection configuration, a bottom-quartile member copies the weights and augmentation hyperparameters of a top-quartile member, then perturbs those hyperparameters and continues training without reinitializing the copied weights. Each policy has two entries for each of 15 operation types, excluding SamplePairing, with one magnitude and one application probability per entry: 60 scalar hyperparameters. The probability grid is 0 to 1 in tenths and magnitude indices are 0 through 9. To augment an image, shuffle the 30 entries, draw an operation-count cap of 0,1 or 2 with probabilities 0.2,0.3 or 0.5, then scan entries and probabilistically apply each until that many operations have actually been applied or the list is exhausted. Each hyperparameter is reset uniformly with probability 0.2; otherwise its integer grid index is shifted up or down by a uniformly sampled amount in {0,1,2,3} and clipped to its domain. The source uses an epoch per step, an exploit/explore interval of three epochs and initially zero probabilities/magnitudes. Retain the selected schedule along the winning training lineage and discard the child weights when training the final model; schedules can be stretched to a different training duration. These numerical settings are the published configuration, not universal PBT requirements.

The pinned exploration code stores probability indices as integers0..10 and magnitude indices as integers0..9, applies reset or signed local perturbations and clips to the matching bounds. The augmentation function decrements its sampled cap only after an operation passes its probability test, stopping at zero or the end of the shuffled list. Repeated operation types are possible because the policy contains two entries per type; applying the same type twice does not require sampling with replacement from a one-entry-per-type list.