Shashankss1205/ML4SCI

This is my personal repository to submit my solutions to the test problems provided by ML4SCI as a selection test for GSOC'24.

ML4SCI

This is my personal repository to submit my solutions to the test problems provided by ML4SCI as a selection test for GSOC'24.

Project Details

Title	Layout	Project	Description
Multi-Class Classification (Common Test 1)	gsoc_proposal	DEEPLENSE	Implementing multi-class classification.
DDPM (Specific Test 4)	gsoc_proposal	DEEPLENSE	Implementing denoising diffusion probabilistic model. [This will be my GSOC'24 proposal project]
SSL on Real Dataset	gsoc_proposal	DEEPLENSE	Implementing SSL Contrastive learning

Diffusion Model

Image generated at first epoch Image generated at 20th epoch

Multi-Class Classification: Common Test 1

Model Architecture	Number of epochs	Learning rate	Training Loss	Training Accuracy	Micro-average ROC-AUC	Macro-average ROC-AUC
ResNet18	50	1e-4	0.00114	1	1.0	1.0
CNN	50	1e-4	1.03	0.461	0.6626	0.64877

ROC-AUC Curve for ResNet18:

SSL on Real Dataset: Specific Test 6

Model Architecture	Number of epochs	Learning rate	Training Loss	Training Accuracy	Lensed AUC	Non Lensed ROC-AUC
Supervised Learning	20	1e-4	0.7248	0.7083	0.9	0.9
SSL with Rotation Augmentation	20	1e-4	0.7336	0.6369	0.82	0.73
SSL with Gaussian Augmentation	20	1e-4	0.6209	0.7619	0.91	0.91

AUC for SSL:

Rotational Augmentation

Gaussian Augmentation

Work done Post GSOC Application Deadline:

Diffusion Denoising Implicit Model (DDIM)

Implemented a DDIM as a part of the Specific Test 4. 2 images generated from that model are shown below: