README.txt

Read Me for CNN-GSGD

Paper: A Strategic Weight Refinement Maneuver for Convolutional Neural Networks, IJCNN 2021 (TBP)
Programmer: Patrick Sharma
Supervisor: Dr. Anuraganand Sharma

The CNN-GSGD code has been written and tested with MATLAB-R2018a. 
Disclaimer: Ensure that you have the nnet toolbox installed under your MATLAB license.

Steps to setup CNN-GSGDon MATLAB-R2018a
You will find the following files in the installation folder:
	- GSGD_setup.bat
	- GSGD_remove.bat
	- 'code' folder
	- 'CNN_Main.m'
	
The 'code' folder consists od the 'CNN-GSGD' code, and all the other necessary code changes needed for 
ease of use - such as enabling/displaying GSGD(isGuided) from the neural network training options.

To setup GSGD on your MATLAB:
	1. Run the 'GSGD_setup.bat' file.
	2. Browse and Select the 'CNN-GSGD' folder with the source code.	
	3. Select the destination 'nnet' folder in MATLAB toolboxes.
		e.g 'C:\Program Files\MATLAB\R2018a\toolbox\nnet'

	4. This will automatically rename the original 'cnn' folder to 'cnn_original' to keep the 
	original matlab cnn files untouched.

	5. You may then add the following additional configurations in the network trainingOptions. 
	   Refer to the sample CNN_Main file
		Set 'isGuided' parameters to true to enalble GSGD and then supply 'Rho', 'RevisitBatchNum' and 					
		'VerificationSetNum' values																								
		Simply remove the above parameters to run without GSGD or set 'isGuided' to false								
																														
			'Rho' - number of iterations  to run for collection and checking of consistent data 
			before guided approach is activated to update the weights with consistent data																		
			'RevisitBatchNum' - number of previous batches to revisit and 
			check how it performs on present batch weights												
																														
			'VerificationSetNum' - number of batches to set aside at the beginning of each epoch.
			Each batch gets picked randomly from this set to attain true error on weights updated by each 
			batch during btraining										

	6. Setup training & Testing dataset and run the 'CNN_Main.m' file to begin training

Note: The actual CNN-GSGD code has been incorporated and can be viewed in the './toolbox/nnet/cnn/+nnet/+internal/Trainer.m' file 
from Lines 72 to 327.

To remove GSGD on your MATLAB:
	1. Run the 'GSGD_remove.bat' file.
	2. Select the 'nnet' folder in MATLAB toolboxes which contains the CNN-GSGD code.
		e.g 'C:\Program Files\MATLAB\R2018a\toolbox\nnet'

	3. This will automatically remove 'cnn' folder with GSGD, and restore the original 'cnn' folder 
	back into matlab.