The code is an adaptation of the Official Fast AutoAugment implementation in PyTorch available here.

Fast AutoAugment is a simplified version of the well known AutoAugment paper of Google Brain.

• Fast AutoAugment learns augmentation policies using a more efficient search strategy based on density matching.
• Fast AutoAugment speeds up the search time by orders of magnitude while maintaining the comparable performances.

The code is also modified so that it can use the augmentation schedule policiy obtained via the official implementation of the Population Based Augmentation available here. Population Base Augmentation is a method that learns a data-augmentation schedule policy and applied during training. The underlying intuition being that at the beginning, there is no need to perform a very elaborated data-augmentation. However, the more the neural network learns, the better the data-augmentation should be. Data-augmentation schedule policies can be found in the folder schedules.

Download CIFAR-100 dataset.

cd /usr/share/bind_mount/scripts/MicroNet/Training
pip install -r requirements.txt
export PYTHONPATH="$PYTHONPATH:/usr/share/bind_mount/scripts/MicroNet/Training"
cd FastAutoAugment
horovodrun -np 4 --cache-capacity 2048 python train.py -c confs/pyramid272a200b.yaml --tag pyramidnet_cifar100 --horovod

The argument np allows you to choose the number of GPUs you want to use (here 4).

Please see the confs/pyramid272a200b.yaml file to get more insights on the parameters used for the training process. You may notice an argument named aug. It allows you to choose which data augmentation policy you want to use.

The checkpoint weights are regularly saved (every 10 epochs) and put in the directory /app/results/checkpoints.