Color Constancy
The aim here was to train a model to learn and predict illumination values from an image.
Given the illumination values, and an image it becomes a trivial task to develop the same image in natural light. This is useful in applicatiosn where images are taken in bad lighting, at night, or indoors in dimly lit places.
In this project I explored the effects of training the model with images in log space.
This is an example of an image taken in the dark.
This is what the image will look like under natural lighting.
I used GoogleNet and MobileNetV3 models for this project.
Feature engineering involved converting images from srgb(linear) to log values.
Data augmentation invovled included arbirary noise in images in training dataset.
Concepts/Packages involved: OpenCV, Neural Networks, Fine Tuning, Feature Engineering, Data Augmentation
Retail Stock prediction
The goal in this project is to build a pipeline, and productionize a model. Used a retail stock prediction dataset, and trained a SGD regressor, KNN regressor and a decision tree regressor to predict howmuch stock movement will occur. Built a datapipeline to train model, incorporated retraining and model selection in the pipleine. Logs, versions of model was written to google cloud.Setup slack alerts so every failure/sucess in the pipeline has a alert sent out.
DAGs shown here perfom Data processing, model training, model selection, logging and staging.
Tools/Concepts involved :Docker, Apache Airflow, ML Flow, Logstash, Google Cloud Platform, MLOps
Assistive writing tool
Played around with OpenAi text model Davinci. Created a simple page where the model response to text prompts are shown, i’ve served the model in my local, deployed it using Flask.
Tools/Concepts involved :Flask, API, Large Language model