Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/facebookresearch/dcem

The Differentiable Cross-Entropy Method
https://github.com/facebookresearch/dcem

Last synced: 8 days ago
JSON representation

The Differentiable Cross-Entropy Method

Awesome Lists containing this project

README 

        
# The Differentiable Cross-Entropy Method



This repository is by

[Brandon Amos](http://bamos.github.io)

and

[Denis Yarats](https://cs.nyu.edu/~dy1042/)

and contains the PyTorch library and source code to reproduce the

experiments in our ICML 2020 paper on

[The Differentiable Cross-Entropy Method](https://arxiv.org/abs/1909.12830).

This repository depends on the

[Limited Multi-Label Projection Layer](https://github.com/locuslab/lml).

Our code provides an implementation of the vanilla

[cross-entropy method](http://web.mit.edu/6.454/www/www_fall_2003/gew/CEtutorial.pdf)

for optimization and our differentiable extension.

The core library source code is in

[dcem/](https://github.com/facebookresearch/dcem/tree/main/dcem);

our experiments are in

[exp/](https://github.com/facebookresearch/dcem/tree/main/exps),

including the [regression notebook](https://github.com/facebookresearch/dcem/blob/main/exps/regression-analysis.ipynb)

and the [action embedding notebook](https://github.com/facebookresearch/dcem/blob/main/exps/cartpole_emb-analysis.ipynb)

that produced most of the plots in our paper;

basic usage examples of our code that

are not published in our paper are in

[examples.ipynb](https://github.com/facebookresearch/dcem/blob/main/examples.ipynb);

our slides are available here in

[pptx](https://github.com/facebookresearch/dcem/blob/main/slides.pptx)

and

[pdf](https://github.com/facebookresearch/dcem/blob/main/slides.pdf)

formats;

and the full LaTeX source code for our paper is in

[paper/](https://github.com/facebookresearch/dcem/tree/main/paper).



![](./fig/dcem.png)



# Setup



Once you have PyTorch setup, you can install our core code as

a package with pip:



```bash

pip install git+git://github.com/facebookresearch/dcem.git

```



This should automatically install the

[Limited Multi-Label Projection Layer](https://github.com/locuslab/lml)

dependency.



# Basic usage

Our core cross-entropy method implementation with the differentiable extension

is available in

[dcem](https://github.com/facebookresearch/dcem/blob/main/dcem/dcem.py).

We provide a lightweight wrapper for using CEM and DCEM in the control

setting in

[dcem_ctrl](https://github.com/facebookresearch/dcem/blob/main/dcem/dcem_ctrl.py).

These can be imported as:



```python

from dcem import dcem, dcem_ctrl

```



The interface for DCEM is:



```python

dcem(

    f, # Objective to optimize

    nx, # Number of dimensions to optimize over

    n_batch, # Number of elements in the batch

    init_mu, # Initial mean

    init_sigma, # Initial variance

    n_sample, # Number of samples CEM uses in each iteration

    n_elite, # Number of elite CEM candidates in each iteration

    n_iter, # Number of CEM iterations

    temp, # DCEM temperature parameter, set to None for vanilla CEM

    iter_cb, # Iteration callback

)

```



And our control interface is:

```python

dcem_ctrl(

    obs=obs, # Initial state

    plan_horizon, # Planning horizon for the control problem

    init_mu, # Initial control sequence mean, warm-starting can be done here

    init_sigma, # Initial variance around the control sequence

    n_sample, # Number of samples CEM uses in each iteration

    n_elite, # Number of elite CEM candidates in each iteration

    n_iter, # Number of CEM iterations

    n_ctrl, # Number of control dimensions

    lb, # Lower-bound of the control signal

    ub, # Upper-bound of the control signal

    temp, # DCEM temperature parameter, set to None for vanilla CEM

    rollout_cost, # Function that returns the cost of rollout out a control sequence

    iter_cb, # CEM iteration callback

)

```



## Simple examples

[examples.ipynb](https://github.com/facebookresearch/dcem/blob/main/examples.ipynb)

provides a light introduction for using our interface for

simple optimization and control problems.



### 2d optimization



We first show how to use DCEM to

optimize a 2-dimensional objective:



![](./fig/2d-iterates.png)



Next we *parameterize* that objective and show how DCEM

can update the objective to move the minimum to a

desired location:



![](./fig/learning.gif)



### Pendulum control



We show how to use CEM to solve a pendulum control problem,

which can be made differentiable by setting a non-zero temperature

for the soft top-k operation.



![](./fig/pendulum-plots.png)

![](./fig/pendulum.gif)



# Reproducing our experimental results

We provide the source code for our cartpole and regression experiments

in the [exps](https://github.com/facebookresearch/dcem/tree/main/exps)

directory.

We do not have plans to open source our PlaNet and PPO experiment.

One starting point is to use an existing PyTorch PlaNet implementation

such as

[cross32768/PlaNet_PyTorch](https://github.com/cross32768/PlaNet_PyTorch)

with a PyTorch PPO implementation such as

[ikostrikov/pytorch-a2c-ppo-acktr-gai](https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail)

or SAC implementation such as

[denisyarats/pytorch_sac](https://github.com/denisyarats/pytorch_sac).



## 1D energy-based regression

The base experimental code for our 1D energy-based regression experiment

is in

[regression.py](https://github.com/facebookresearch/dcem/blob/main/exps/regression.py).

Once running this, the results can be analyzed with

[regression-analysis.ipynb](https://github.com/facebookresearch/dcem/blob/main/exps/regression-analysis.ipynb),

which will produce:



![](./fig/regression-loss.png)

![](./fig/regression-surface.png)

![](./fig/regression-iter-ablation.png)

![](./fig/regression-iter-ablation-2.png)



## Embedding actions in the cartpole

The base experimental code for our cartpole action embedding

experiment is in

[cartpole_emb.py](https://github.com/facebookresearch/dcem/blob/main/exps/cartpole_emb.py).

Once running this, the results can be analyzed with

[cartpole_emb-analysis.ipynb](https://github.com/facebookresearch/dcem/blob/main/exps/cartpole_emb-analysis.ipynb),

which will produce:



![](./fig/cartpole-cem-iter-vis.png)

![](./fig/cartpole-cem-iter-vis-latent.png)



# Citations

If you find this repository helpful in your publications,

please consider citing our paper.



```

@inproceedings{amos2020differentiable,

  title={{The Differentiable Cross-Entropy Method}},

  author={Brandon Amos and Denis Yarats},

  booktitle={ICML},

  year={2020}

}

```



# Licensing

This repository is licensed under the

[CC BY-NC 4.0 License](https://creativecommons.org/licenses/by-nc/4.0/).