Hello!
I'm Krut Patel;






About Me
# recent visual computing graduate
# machine learning and deep learning enthusiast
# interested in computer vision, NLP and Recommendation Systems
# past freelance software and game developer

Download CV Blog

Background

Work Experience

Projects

Computer Vision and Graphics

Convolutional Siamese Network for Writer Independent Offline Signature Verification

# Base Model - SigNet model with contrastive loss

# Model variation 1 - using triplet loss

# Model variation 2 - using binary cross entropy loss where the network outputs the L1 component wise distance between feature vectors outputted by each Siamese twin

# Datasets : CEDAR, GPDS300, GPDS Synthetic SignatureDatabase and BHSig260

Technologies used - PyTorch, TorchVision, Pillow, Matplot, Numpy

Check it out on GitHub

MNIST 3D voxelized Digit Generation using GAN

# 3D Generative Adversarial Network

# Generate 3D hand-written digits, represented in voxels

# Latent space Interpolation that allows interpolating between two voxelized digits

Technologies used - PyTorch, TorchVision, OpenCV, Matplot, Numpy

Check it out on GitHub

Image Inpainting using UNet and ResNet

# Reconstructing lost or deteriorated parts of images using CNN

# Data Augmentation

# UNet and ResNet models

# MSE loss function

Technologies used - PyTorch, OpenCV, Matplot, Numpy

Check it out on GitHub

Image Segmentation using UNet

# Locate objects and boundaries in images

# Using UNet model

# Batch Normalization and Data Augmentation

# Implementation for Cross entropy loss function and softmax

# MSE loss function

Technologies used - PyTorch, OpenCV, Matplot, Numpy

Check it out on GitHub

Quadric based Mesh Decimation

# Implementation of Winged Edge Structure to save 3d model

# Calculating Flat and Smooth Normals

# Quadric Error Calculation

# Mesh Simplification with Edge Collapse

# Saving model from Winged Edge to .obj

Technologies used - C++, OpenGL, NanoGUI

Check it out on GitHub

Vehicle Detection

# Vehicle detection and segmentation

# SSD (Single Shot MultiBox Detector), Mobilenet v2 as BaseNet

# Datasets: Cityscapes

Technologies used - PyTorch, OpenCV, Matplot, Numpy

Facial Landmark Regression

# SSD Model with Alexnet as backbone network

# Data augmentation and transfer learning

# Dataset: Labeled Faces in the Wild(LFW)

Technologies used - PyTorch, OpenCV, Matplot, Numpy

ICP SLAM for safe robot teleportation

# Localization and Mapping for a Turtlebot robot using LIDAR sensor

# ICP SLAM Algorithm

Technologies used - C++, ROS Framework

MNIST 2D GAN

# 2D Generative Adversarial Network

# Generating hand-written digits

Technologies used - PyTorch, TorchVision, OpenCV, Matplot, Numpy

Check it out on GitHub

MNIST Handwritten Digits Classification using a Convolutional Neural Network (CNN)

# Classifying MNIST digits using a convolutional neural network implemented in PyTorch

# Developed as a part of a Medium blog post explaining CNN concepts

Technologies used - PyTorch, TorchVision, Matplot, Numpy

Blog post Medium
Check it out on GitHub



Natural Language Processing

Toxic Comment Classification

# Kaggle competition - Multi-label sentence classification

# Model 1: Logistic Regression using TF-IDF

# Model 2: Stacked Bidirectional LSTM

# Model 3: CNN by Yoon Kim

# Using pretrained word embeddings

Technologies used - PyTorch, Numpy, Keras, Seaborn, Matplotlib


Check it out on GitHub

Robust Phrasal Chunking

# Finding non-recursive syntactic groups of words

# Using semi-characterRNN + LSTM to predict the phrasal chunking tags

# Concatenated semi-characterRNN's hidden state with word embeddings as input to LSTM

# Model learns word embeddings to minimize the loss on the phrasal chunking task

# Dataset : CoNLL 2000 shared task2

Technologies used - PyTorch, Numpy

Lexical Substitution with Retrofitting

# Finding a suitable substitution for a target word in a sentence

# Using WordNet ontology for semantic lexicon

# Retrofitting word2Vec word vectors with semantic lexicons

# Incorporating Context Words around the target word to find better substitute words

# Dataset : SemEval-2007 Task 10: English Lexical Substitution Task

Technologies used - Python, Numpy
Check it out on GitHub

ZH Chinese Text Segmentation

# Segmenting a sequence of Chinese characters into the most likely word sequence

# Segmentation scored based on the probability of the words that occur in that segmentation

# Unigram, Bigram and Trigram Model

# Jelinek Mercer smoothing and Stupid Backoff smoothing

Technologies used - Python, Numpy
Check it out on GitHub

EN English Text Segmentation

# Segmenting a sequence of English characters into the most likely word sequence

# Segmentation scored based on the probability of the words that occur in that segmentation

# Unigram model with smoothing for unknown unigrams

Technologies used - Python, Numpy
Check it out on GitHub

Quora insincere questions classification - using CNN

# Yoon Kim's Sentence classification CNN model

# Binary text classification with imbalanced classes

# Comparing CNN with Traditional Models (TFIDF + Logistic Regression and SVM)

# Predicting if a question on Quora is sincere or not

# Datasets : Dataset - Quora questions from a Kaggle competition

Technologies used - PyTorch, TorchText, ScikitLearn, Matplot, Numpy



Time Series Forecasting

Predict Future Sales

# Kaggle competition - Time series problem

# Dataset with two years of item sales count for various stores across Russia

# Predicting item sales for the next month

# Exploratory Data Analysis, Clustering and Feature Extraction

# Using Stacked LSTM, XGBoost and LSTM Autoencoder

Technologies used - PyTorch, Numpy, Keras, Scikit-Learn, Plotly, Matplotlib


Check it out on GitHub



Game Prototypes


Timberman

Clone of game Timberman.

Check it out on GitHub


Knife Frenzy

Clone of game Knife Hit.

Check it out on GitHub


Nail It

Clone of game Nail It.

Check it out on GitHub


Tappy Dunk

Clone of game Tappy Dunk.

Check it out on GitHub


Cricket Prototype

Prototype for mobile cricket game.

Check it out on GitHub


Maze AR

Find your way through a augmented reality maze.

Coming soon on Google Play.


Arcadia

Swipe your finger to guide the player to avoid the swinging rings.

Coming soon on Google Play.


Round Pong

It's pong but in a circular platform and played by a single player.

Coming soon on Google Play



Published Games


Spin Out

Find your way dashing through rings.

Get it on Google Play.


Retro Skies

Mayday! Mayday! Avoid the missiles to survive.

Get it on Google Play.


Dodge

Swipe your finger to guide the player to dodge the falling blocks.

Get it on Google Play.


Adventure Heroes

Hop the character by tapping left or right part of the screen and avoid obstacles.

Get it on Google Play.


Morph

Tap to morph shape and color of player to match that of obstacles.

Get it on Google Play.

Get In Touch

Don't hesitate to contact me for freelance contracts or to just say 'Hello'!