Hi there, I am Himali Saini
MTech in Computer Science at IIT Bombay
● Implemented advanced deep learning techniques such as U-Net, ResNet, GANs, and variational Auto-Encoders to achieve
high-quality super-resolution for 4x and 8x upscaling tasks.
● Developed and applied custom perceptual loss metrics to train and evaluate models, resulting in a significant
improvement of more than 2 times in perceptual similarity scores.
The above sample images have been tested on 8x low resolution blurred test set, it shows the model outputs for EDSR 8x and SRGAN 8x.
SRGAN has been implemented using EDSR as the generator, we have used a custom loss function for the generator, a combination of SSIM, LPIPS and L1 losses.
● Implemented Neural Style Transfer using two different deep learning techniques.
● CNN: Extracted feature maps using VGG19, and plotted loss curves for various subsets of feature maps to analyze
content and style reconstruction. Used parameter tuning to find the best style and content-weighted image.
● CycleGAN: Learnt to map between input and output images using unpaired datasets (style and content images).
Have a look at the above gif, adding the magic of Van Gogh's Starry night to a unicorn image!
● Implemented insert and update violation checks for functional dependencies in Postgres utilizing the SPI library
● For every relation R we store the set of functional dependencies as FD (in the form of a relation), for every functional dependency F (Assume X -> Y) belonging to this set, whenever an insert or update is performed there is no violation for F, i.e for every value of X there should be only one corresponding value of Y. Since we are checking for every insertion, every database instance will follow the FD rule.
● Scraped product information for trending electronic items for a period of one month and performed data visualization.
● Created a web app to search for products to show prices and review comparisons between Amazon and Flipkart.
● Added history and trending pages, making use of Sqlite database.
● Implemented and trained word2vec models from scratch for extracting word embeddings for the analogy task.
● Implemented encoder-decoder architecture using bidirectional LSTM for POS tagging task, with an accuracy of 98%.
Django based website for a cafe. Users can add dishes to cart, enter billing details ,place order and download receipt in pdf format. Also users can post , update and delete reviews and maintain their profile details along with log in and sign up functionality.
Flutter based app where users can browse for various songs, artists, add to favourites, with basic authentication features.
