toy_gradlogp_tf2

This repo implements some toy examples of the following score matching algorithms in TensorFlow 2.0:

• ssm-vr: sliced score matching with variance reduction
• ssm: sliced score matching
• deen: deep energy estimator networks
• dsm: denoisnig score matching

Related projects:

• toy_gradlogp: PyTorch Implementation.

Installation

Basic requirements:

• Python >= 3.6
• TensorFlow >= 2.3.0

Install from PyPI

pip install toy_gradlogp_tf2

Or install the latest version from this repo

pip install git+https://github.com.Ending2015a/toy_gradlogp_tf2.git@master

Examples

The examples are placed in toy_gradlogp/run/

Train an energy model

Run ssm-vr on 2spirals dataset

python -m toy_gradlogp.run.train_energy --loss ssm-vr --data 2spirals

To see the full options, type --help command:

python -m toy_gradlogp.run.train_energy --help

usage: train_energy.py [-h] [--logdir LOGDIR]
                       [--data {8gaussians,2spirals,checkerboard,rings}]
                       [--loss {ssm-vr,ssm,deen,dsm}]
                       [--noise {radermacher,sphere,gaussian}] [--lr LR]
                       [--size SIZE] [--eval_size EVAL_SIZE]
                       [--batch_size BATCH_SIZE] [--n_epochs N_EPOCHS]
                       [--n_slices N_SLICES] [--n_steps N_STEPS] [--eps EPS]
                       [--log_freq LOG_FREQ] [--eval_freq EVAL_FREQ]
                       [--vis_freq VIS_FREQ]

optional arguments:
  -h, --help            show this help message and exit
  --logdir LOGDIR
  --data {8gaussians,2spirals,checkerboard,rings}
                        dataset
  --loss {ssm-vr,ssm,deen,dsm}
                        loss type
  --noise {radermacher,sphere,gaussian}
                        noise type
  --lr LR               learning rate
  --size SIZE           dataset size
  --eval_size EVAL_SIZE
                        dataset size for evaluation
  --batch_size BATCH_SIZE
                        training batch size
  --n_epochs N_EPOCHS   number of epochs to train
  --n_slices N_SLICES   number of slices for sliced score matching
  --n_steps N_STEPS     number of steps for langevin dynamics
  --eps EPS             noise scale for langevin dynamics
  --log_freq LOG_FREQ   logging frequency (unit: epoch)
  --eval_freq EVAL_FREQ
                        evaluation frequency (unit: epoch)
  --vis_freq VIS_FREQ   visualization frequency (unit: epoch)

Results

Tips: The larger density has a lower energy!

8gaussians

Algorithm	Results
ssm-vr
ssm
deen
dsm

2spirals

Algorithm	Results
ssm-vr
ssm
deen
dsm

checkerboard

Algorithm	Results
ssm-vr
ssm
deen
dsm

rings

Algorithm	Results
ssm-vr
ssm
deen
dsm