PyTorch implementation of Gapped Straight-Through Estimator (GST) with experiments on MNIST-VAE and ListOps. We compare our proposed GST estimator with several discrete random variable estimators including Straight-Through Gumbel-Softmax (STGS) and Rao-Blackwellized Straight-Through Gumbel-Softmax (rao_gumbel).
- Gapped Straight-Through Estimator (GST): Training Discrete Deep Generative Models via Gapped Straight-Through Estimator [ICML 2022].
- Straight-Through Gumbel-Softmax (STGS): Categorical Reparametrization with Gumbel-Softmax
- Rao-Blackwellized Straight-Through Gumbel-Softmax (rao_gumbel): Rao-Blackwellizing the Straight-Through Gumbel-Softmax Gradient Estimator
- MNIST-VAE experiment is based on YongfeiYan's implementation.
- ListOps experiment is based on Learning to Compose Task-Specific Tree Structures and its implementation. The dataset comes from ListOps: A Diagnostic Dataset for Latent Tree Learning.
We recommend using Anaconda with the following commands:
conda create -n GST python=3.8
conda activate GST
conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt
--mode (default: gumbel) selects the estimators. Possible choices are gumbel, rao_gumbel, gst-1.0, and gst-p.
--temperature (default: 1.0) controls the temperature of the softmax function for the soft sample.
--hard (default: True) gives hard samples using the straight-through trick; otherwise, soft samples are generated.
Example 1: train STGS at temperature 1.0
python gumbel_softmax_vae.py --mode gumbel --temperature 1.0
Example 2: train GST-1.0 at temperature 0.5
python gumbel_softmax_vae.py --mode gst-1.0 --temperature 0.5
Example (a): train rao_gumbel at temperature 0.1
python -m nlp.train --word-dim 300 --hidden-dim 300 --clf-hidden-dim 300 --clf-num-layers 1 --batch-size 16 --max-epoch 20 --save-dir ./checkpoint_listops --device cuda --pretrained glove.840B.300d --leaf-rnn --dropout 0.5 --lower --mode rao_gumbel --task listops --temperature 0.1
Example (b): train GST-1.0 at temperature 0.1
python -m nlp.train --word-dim 300 --hidden-dim 300 --clf-hidden-dim 300 --clf-num-layers 1 --batch-size 16 --max-epoch 20 --save-dir ./checkpoint_listops --device cuda --pretrained glove.840B.300d --leaf-rnn --dropout 0.5 --lower --mode gst-1.0 --task listops --temperature 0.1