README.md

In dfalbel/JEM: An implementation of Joint Energy-based Model.

JEM

An R implementation of ‘Your Classifier is Secretly an Energy Based Model and You Should Treat it Like One’. Code heavily inspired by https://github.com/tohmae/pytorch-jem and https://github.com/wgrathwohl/JEM.

Installation

JEM is not on CRAN and can be installed with:

remotes::install_github("dfalbel/JEM")

Experiment

You can run an experiment with the run_experiment function. The configuration is found by default in the config.yaml file in the root directory.

experiment <- JEM::run_experiment()
#> 
#> [36m──[39m [1m[1mExperiment config[1m[22m [36m────────────────────────────────────────────────────────────────────────────────────────────────[39m
#> → dataset : mnist
#> → model : mlp
#> → n_epochs : 5
#> → lr : 0.001
#> → rho : 0.05
#> → eta : 20
#> → sigma : 0.01
#> → alpha : 1
#> → buffer_size : 10000
#> → device : cuda
#> → batch_size : 100
#> → conditional : TRUE
#> ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
#> [Epoch 1] Train{Loss: 0.856000} Valid{Loss: 0.258047, Acc: 0.926600}
#> [Epoch 2] Train{Loss: 0.131079} Valid{Loss: 0.147414, Acc: 0.960800}
#> [Epoch 3] Train{Loss: 0.178203} Valid{Loss: 0.126619, Acc: 0.964900}
#> [Epoch 4] Train{Loss: 0.160592} Valid{Loss: 0.124880, Acc: 0.961600}
#> [Epoch 5] Train{Loss: 0.138205} Valid{Loss: 0.093921, Acc: 0.971300}

After the model is run, you can generate samples from (p(x|y)) with:

samples <- JEM::generate_samples(
  experiment, 
  rep(1:9, each = 10), 
  eta = 200
)
JEM::plot_samples(samples, nrow = 10)

dfalbel/JEM documentation built on Oct. 26, 2020, 2:44 p.m.