Color-Spatial-AutoAugment

The Implementation used to test Color-Spatial AutoAugment in "Color-Spatial AutoAugment Another Approach To AutoAugment Policies Found & Implementation".

Pretrained Models and their TensorBoard can be found Here.

Required Packages

• Torch Package.
• TorchSummary
• tqdm
• scipy
• scikit-learn
• seaborn
• tensorboard

install the required packages using

pip install -r requirements.txt

Or run venv_Install.ps1 to create the venv and install the required packages.

Training Steps

• Run the venv or run venv_Run.ps1 to activate the venv.
• run main.py to start the training or edit config.py for extra settings.
• --augment have two options AutoAugment to train using AutoAugment and Augment to run using Color-Spatial AutoAugment.
• --dataset either cifar10 , cifar10_reduced orcifar100.
• --network either resnet18 , resnet50 and wideresnet for WideResNet28-10

After running the required epochs, run main_linear.py for the learner classifier.

and example of training using Color-Spatial AutoAugment with backbone ResNet50 for 1000 epochs

python main.py --augment Augment --dataset cifar10 --network resnet50 --epochs 1000 --batch_size 64

• Important Note: All tests were done on batch size 64 due to memory restrictions on the testing setup.

The table shows the accuracy of Cifar-10 with both AutoAugment and Color-Spatial AutoAugment on the different networks; for the ResNet50, we Show the accuracy for both 300 epochs and 1000.

Epochs	Network	AA	CS-AA
	ResNet18	77.48	82.49
300	ResNet50	79.51	85.74
	WideResNet 28-10	79.01	82.13
1000	ResNet50	85.56	91.1