Data augmentations have been widely studied to improve the accuracy and robustness of classifiers.

However, the potential of image augmentation in improving GAN models for image synthesis has not been thoroughly investigated in previous studies.

In this work, we followed the systematically study of the "Image Augmentations for GAN Training" writers,

and recover some of their results, while we implemented more new augmentation techniques for GAN training in a variety of settings.

We validate the claim that augmentation for real and fake images before Discriminator phase, does improve the learning while.

We also validate the claim that augmentations that result in spatial changes improve the GAN performance more than those that induce mostly visual changes.

We followed the systematically analysis method for evaluating the results using FID measurement.

Our analysis was built from combining many pieces of codes, and using new code for augmentation implementation, and for experiments and analysis.

1. Structure of the code:

    folders hyrarchy:
   
    ./src/					- main source code (explanation next)
   
    ./src/data/				- CIFAR10 dataset place holder (automatically to be downloaded from web)
   
    ./src/FID/				- FID calculation files
   
    ./src/weigts/			- weigts place holder (automatically to be created when run)
   
    ./src/outputs/			- outputs place holder (automatically to be created when run)
   
   
   
    1.1 ./src/FID/ : FID Calculaion files.
   
    	1.1.1 fid_score.py - fid score calculation, operates using calculate_fid_given_paths function
   
   
    	      main functions:
   
    	      def getInceptionModel			- prepare for FID calc
   
    	      def getRealm1s1				- prepare for FID calc
   
    	      def calculate_fid_given_paths		- FID calc
   
    	1.1.2 inception.py - generation of InceptionV3 model used as part of fid score calculation
    
   
    1.2 ./src/FID/Augmentation.py : augmentation file. 
    	implementation of translationX, translationY, colorNoise
   
    	def Augmentation(I,lam,aug_type='translationX',device='cuda')
    	
   
   
    1.3 ./src/FID/util.py : utility functions.
   
    	def load_weights(netG, path ='weights/netG_epoch_24.pth',device='cpu'):
   
    	def createDir(aug_type, lam)
   
    	def train_loop(..) - the training loop over batches and epoches for training of Generator and Discriminator 
   
    	with augmentation.
    	
   
    1.4 ./src/FID/test.py : experiment file process:
   
    	# %% Imports
   
    	# %% Configs.
   
    	# %% Generator
   
    	# %% Discriminator
   
    	# %% optimizer & Loss Config
   
    	# %% Train Loop (over aug_types , lambda vals): aug_type_vec   = ['colorNoise','gammaCor','translationY'], lambda_vec = np.arange(0.1, 1.1, 0.1).round(2)
   
    	# %% create folders - load wights (for aug type and lmbda val) and produce directory with results.
   
    	# %% prepare FID - prepare Inception model and m1, s1 value based on real data
   
    	# %% calc FID per dir 
    	
   
   
    1.5 py-sbatch.sh, jupyter-lab.sh  - file to create job process to run in the CS servers.
    
   
    1.6 gan_cifar.py, gan_cifar.ipynb - files for manully further research. not relevant for reproduce resutls.
   

2. steps to reproduce the results:

   2.1 run in tesy.py lines: 165, decide augmentation and lambda values. default - all of them (be carefull it's long!)

   2.2 procedure to run test.py on the servers:

    - conda activate cs236781-hw
    
    - cd FINAL/DCGAN_CIFAR10/
    
    - srun -c 2 --gres=gpu:1 --pty python -m test.py
   

   2.3 if you want to check the net manualy in the jupyter lab, you can use the notebook gan_cifar.ipynb

    - srun -c 2 --gres=gpu:1 --pty jupyter-lab.sh