I made a sentiment analysis model using a transformer architecture from scratch. Instead of using a pre-trained BERT model I used last time, I decided to make my own with BERT. It was actually very difficult as the math and library management was not working properly. I was running into file issues all the time as well. I was using IMDb dataset, which I have used previously which helped my efficiency.
github.com/Shahrishi4324/Transformer-based-sentiment-analysis
I created a deep learning neural network that integrates CNN and RNN to generate captions for images. The model is trained on image-caption pairs and is capable of generating descriptive captions for new images.
github.com/Shahrishi4324/ImageCaptioning
I implemented a transformer model from scratch using PyTorch, trained it on the Multi30k dataset for machine translation from German to English, and created a transformer that is a model that can leverage self-attention mechanisms to achieve high performance in sequence-to-sequence tasks. Here is the github repository.
github.com/Shahrishi4324/Machine-translation-with-transformer
So I made a gesture controlled robotic arm. Using OpenCV, it tracks the gestures you make with your fingers to the laptop, converts it into 0 and 1s like how 0 means finger is up, and 1 is finger is down, and then sends it to the serial monitor where the arduino code converts that into instructions for the robotic arm. The issue lies within the robotic arm as of right now, the servos i am using dont have enough torque and i need to purchase some new ones. However the setup is all good for the robotic arm, so after I switch in the new servos, it will be able to run perfectly.
github.com/RishiShah99/GestureControlledRoboticArm
I created my own voice virtual assistant called Zephyr. It detects my voice, and is a windows service on my machine. I am still developing it to be much better on my machine with the windows service, but the actual app uses NLP to detect your voice, identify key words, and then can do tasks, like creating events on your calendar, tell you the location in a specific place, has news information depending on whatever topic you want to converse about, and has music playback as it has access to spotify. I am going to 3D print an actual system for Zephyr where it lights up inside the little hub, and i can converse with it via raspberry Pi. I have started development on that part. But as of right now, in the first block, it has access to a bunch of apis. I am also creating a website for it and am going to add keybinds to it.
Here is the github repository for it:
github.com/RishiShah99/Zephyr
It speaks using pyttsx so the stuff printed is just to show see what state it is in right now.
I fine tuned a pre-trained Named Entity Recognition model with SpaCy. I loaded a real-world dataset, annotated the text with named entities. I have really just been working on natural language processing with spacy recently.
github.com/Shahrishi4324/Fine-Tuning-Named-Entity-Recognition-with-SpaCy
I conducted text classification with BERT on the imdb datasets. I did sentiment analysis on the imdb dataset, which has movie reviews labelled as positive or negative. I loaded a pre-trained BERT model and then fine tuned it on the dataset.
github.com/Shahrishi4324/TextClassification
I made a project that involves building a sentiment analysis model to classify movie reviews as positive or negative. I used SpaCy for text preprocessing and Scikit-learn for model training and evaluation.
github.com/Shahrishi4324/MovieReview
Made a CNN using Keras and then performed adversarial attacks on the machine learning model. I was running into issues with processing the images. I used a new adversarial attack called FGSM onto the model as well. github.com/Shahrishi4324/Adversarial-Robustness-with-Keras
I made a 3d maze explorer. It has AI enemies that chase the player through the maze, has collectible items and dangerous traps within the maze, detects collisions with walls.
github.com/Shahrishi4324/3DMazeExplorer/tree/main
I made a game on OpenCV where there is scoring, power ups, obstacles. You move an object based off your hand gestures.
github.com/Shahrishi4324/Open_CVGame
I made an app where it turns on the camera and then you can do makeup onto it with sliders using TKinter and OpenCV.
github.com/Shahrishi4324/MakeUpApp
I made the Website for my Pose Estimation Model that tracks movements from photos and videos. The website only has photo capabiltieis as videos take too long to segment properly. It then shows the segmented image after you upload one.
Created a 2 player game with Pong. It is extremely smooth and two people can play at once.
The Space Game, pretty much your shooting a bunch of aliens that are coming down to break your ship. If you shoot the bombs coming down to hit the ship, a sprite animation loads of the bomb blasting. Additionally, if you shoot the alien, they also blow up with an animation.
Made a store inventory system. You can add any type of item onto the store. Add the stock number for that item. Name the item and the amount of that item. You can discontinue an item, check the low stocks, get the stocks of certain items. Might create a better GUI for this to make it seem like you are managing a store fully. Lots of fun working with arraylists in java.
A game where you catch water droplets with a bucket. Used Java Swing and worked with images.
I made hangman in TKinter and Python. I made a lot of mistakes with creating the lines, and trying to make the hangman on TKinter, every single accurate placement was very hard. It can tell whether you used that letter or not. Ensures that you only places letters as well
Added in Carlini Wagner Attacks onto my Original ResNet18 Machine Learning Model trained on the ImageNet100 dataset. This model was extremely weak against the Carlini wagner attacks.Additionally, i spent an hour trying ot figure out the best way to commit to Github via kaggle but it kept saying that it was too large of a file. Very annoying.
Created a ResNet18 Model for the MNIST dataset. I used PGD attacks and Carlini Wagner Attacks onto the machine learning model.
Made Plants vs Zombies on Java. It has the exact same sprite loading iamges, the exact same music, and perfect animations. It has the currency. And everything.
Created a payroll structure for a company. I might turn this into a gui next.
Fruit Gunner , instead of slicing the fruit, you shoot the fruit.
I made a to-do task on tkinter. This was my hardest project on tkinter so far, but I have finally gotten the hang of it after 4 projects. I think I can start implementing into my big projects now. This can add tasks, delete tasks, shows all tasks, complete the task, save the task as a file, and load tasks from a file as well.
Basic Drawing App on TKinter
Calculator made in tkinter
Created Tic Tac Toe on Python using Tkinter.
Created a pose estimation tracker. It takes in a video and then uses a pose estimation ml model that I developed to find keypoints on the human, and then returns out a new video.
I created a memory game. It flips the cards if you get them wrong, if you get it right, both cards disappear.
I created brick breaker on java. Can you please check it out
Tic tac toe made in java
basic calculator made in java
So I created a resnet18 model for ImageNet 10 and imagenet100. Here is what it looks like:
It is an image classifier model for both. The ImageNet100 was able to achieve a top 5 accuracy of 94%!
I have finished the hand portion to my gesture controlled robotic arm. It has 5 fingers, each that are bendable, and has a wrist that also bends inwards as if you are grasping something. There is a stand that will be connected to the forearm later. This took a lot of CAD and lots of time 3D printing as well. Here is the project:
I have finished my ResNet18 Model on the ImageNet10 Dataset that implements PGD attacks on it as well. The PGD attacks completely broke my machine learning model and it was very fun to use TorchAttacks. Here is my github repository for it. github.com/RishiShah99/imagenet10-resnet18-pgd-attacks/tree/main
Created a croc jibbit that will be 3D printed later. It has the OVO logo on top of it, and the dimensions were all created to fit my crocs perfectly. Here is an image of it
