A greedy approach for finding the optimal architecture (in terms of speed-accuracy tradeoff) for Multi-Task Learning. Although such kind of delicate graph manipulations are easier on PyTorch, we chose to use TensorFlow for its matureness on mobile platforms (TF-Lite).
This repo is divided into 3 main branches:
master
- for the main Unzip-Nets Multi-Task Learning codeprototyping
- for random Jupyter notebooks, small experiments, and prototyping snippetsgh-pages
- for the documentations and literature summaries (with LaTeX)