Deep Learning From Scratch: Vision & Multimodal
Prototype
EfficientNet-B0
Real coefficients and ImageNet results from the EfficientNet paper (Tan & Le, 2019) — the same compound scaling rule re-implemented from scratch in PyTorch. Each column is one of B0's seven MBConv stages: depth scaling stacks more blocks, width scaling fattens their channels, resolution scaling grows the input.
Image Caption Generator via Transformers (Sept 2022)
- Trained an encoder-decoder attention network using the Flickr8k dataset, extracting spatial feature volumes via a softmax-free InceptionV3 network mapped through multi-head self-attention positional blocks, evaluated via BLEU scoring.
EfficientNet Core Re-implementation (June 2022)
- Coded the complete compound model scaling architecture from the foundational EfficientNet: Rethinking Model Scaling for CNNs paper from scratch using PyTorch, writing custom Inverted-Residual blocks scaled exactly to specific resolution expansion coefficients (φ).
Automated Vacuum Robotics (Jan 2021)
- Modeled and tested circuit designs in Tinkercad for an affordable autonomous cleaning robot featuring real-time Bluetooth smartphone interaction, ultrasonic obstacle avoidance paths, and electronic cleaning brushes.
Technologies
- InceptionV3
- PyTorch
- EfficientNet
- Tinkercad
- Bluetooth