The goal of this project was to take an existing Poincare Embedding implementation and try to optimize it. I used Numba, CuPy for GPU performance, and a lot of python vectorization to try and achieve faster speeds.
The datasets were retrieved from Facebook's Poincare Implementation.
An interactive sample where you can hover over points can be found in sample.html
from embed import PoincareEmbedding
embedding = PoincareEmbedding(closure, ...params)
embedding.fit_transform(100)
// Access embedding using embedding.embedding
Running on the mammals dataset with 100 negatives, batch size of 1, and 300 epochs.
Epoch 1 (after 10 epochs of burn in with 1/10 the learning rate)
Epoch 30
Epoch 100
Epoch 200
Epoch 300
