Anton – Python, Machine learning, Keras

In this project, a neural networks were trained to classify the type of car breakdown by sound. This algorithm is needed to replace diagnostics made by people with machine learning. Various models and approaches to data preprocessing were tested.

• data preparation;
• training neural networks;
• metrics selection;
• models evaluation;
• model deployment;

Participated in writing a custom framework for deep learning based on PyTorch Lightning and Hydra.

• Added new NN architectures training pipelines;
• implemented a new functionality.

This project was completely made by Anton from scratch. More than 5 million sports cards, along with metadata, were scraped from the Internet. After this, an artificial dataset was generated for training the card detector. The YOLO detector was trained, and embedding vectors were generated for each card. The final program had 2 parts (API and UI), which could detect cards in photos and find the most similar ones in the database.

• web scraping;
• artificial dataset generation;
• neural networks training for object detection;
• models evaluation;
• embeddings vector generating;
• API deployment;
• creating a graphical interface.

Training of an ML model for the prediction of food item cuisine based on the name, description, and price of restaurant menu items. It was made for HoReCa in the USA. Having a lot of menus from different restaurants (by web scraping), Anton analyzed what type of food is in trend with this model.

• data preparation;
• metrics selection;
• models training;
• models evaluation;
• API deployment.

Monodepth estimation model. This was done for a project in Samsung, which is under the NDA. An approach called "Deeper Depth Prediction with Fully Convolutional Residual Networks" (https://arxiv.org/abs/1606.00373) was used here.

• data preparation;
• models performance evaluation;
• metrics selection;
• neural networks training on GPUs.

Object detector that will work quickly for the video. It was done for a project in Samsung, which is under the NDA. An approach called "Deep Feature Flow for Video Recognition" (https://arxiv.org/abs/1611.07715) was used here.

• An investigation of different deep learning approaches;
• data preparation;
• models performance evaluation;
• neural networks training on GPUs.