Differentiable Discrete Samplers (d2sample)

d2sample is a collection of PyTorch differentiable samplers for discrete objects, with associated examples and layers (TBD). Install by first cloning recursively:

git clone --recursive git@github.com:sscardapane/d2sample.git

Then (for now) add to the path:

import sys
sys.path.append('./d2sample/')

Requirements

• PyTorch == 1.13.1
• functorch

Implemented algorithms

$k$-subset sampling (see notebook):

1. Gumbel-Softmax with continuous top-$k$ relaxation (Xie & Ermon, 2019). For $k=1$ this reduces to the standard Gumbel-Softmax reparameterization available inside PyTorch.
2. Top-k selection with I-MLE (Niepert, Minervini, & Franceschi, 2021).
3. SIMPLE: Subset Implicit Likelihood Estimation (Ahmed, Zeng, Niepert, & Van den Broeck, 2022).