classic ImageNet training pipeline using contemporary best practices in data efficiency, augmentation, and optimization. Targets ≥75% top-1 from scratch within budget constraints.

maxing accuracy on mnist with under 20k paramters, and under 20 epochs of training

Chrome extension to dim, blur, or mask any part of a webpage. Select what matters; focus instantly.

FaceLock is a privacy-first face re-identification demo.

AboutMe

a simple for personal use to keep track of my session

it encoders all the way

just some work on smol model training, optimization, etc...

using semantic segmentation to identify drivable path for self driving cars

Two-stage zero-shot evaluation pipeline for STEM-en-ms dataset

a couple fine tuning jobs

running colab on selenium

template for pl + hydra + timm + cog

use github API to find stale branches

legal maxims dataset

a series of attempts on vision transformers

a mock system to do ocr on document, and index for the search and retrieval

some experiments on learning with noisy data using the clothing1M datset.

DETR (like) architecture for detecting defects in concrete surface

a template for my py38 container, with only basic docker-compose and util scripts