Gated Path Planning Networks (ICML 2018)

This is the official codebase for the following paper:

Lisa Lee*, Emilio Parisotto*, Devendra Singh Chaplot, Eric Xing, Ruslan Salakhutdinov. Gated Path Planning Networks. ICML 2018. https://arxiv.org/abs/1806.06408

Getting Started

You can clone this repo by running:

git clone https://github.com/lileee/gated-path-planning-networks.git
cd gated-path-planning-networks/

All subsequent commands in this README should be run from the top-level directory of this repository (i.e., /path/to/gated-path-planning-networks/).

I. Docker container

We provide two Docker containers, with and without GPU support. These containers have Python 3.6.5, PyTorch 0.4.0, and other dependencies installed. They do not contain this codebase or the maze datasets used in our experiments.

To load the container with GPU support:

# PyTorch with GPU support
nvidia-docker pull lileee/ubuntu-16.04-pytorch-0.4.0-gpu:v1
nvidia-docker run -v $(pwd):/home --rm -ti lileee/ubuntu-16.04-pytorch-0.4.0-gpu:v1

To load the container without GPU support:

# PyTorch (CPU-only)
nvidia-docker pull lileee/ubuntu-16.04-pytorch-0.4.0-cpu:v1
nvidia-docker run -v $(pwd):/home --rm -ti lileee/ubuntu-16.04-pytorch-0.4.0-cpu:v1

Here is a speed comparison between the Docker containers for training VIN on a 9x9 maze with 5k/1k/1k train-val-test split:

PyTorch 0.4.0	time per epoch
with GPU support	8.5 sec
without GPU support	32.3 sec

II. Generate a 2D maze dataset

Generate a dataset by running:

python generate_dataset.py --output-path mazes.npz --mechanism news --maze-size 9 --train-size 5000 --valid-size 1000 --test-size 1000

This will create a datafile mazes.npz containing a dataset of 9x9 mazes using the NEWS maze transition mechanism with 5k/1k/1k train-val-test split.

Note:

The same maze transition mechanism that was used to generate the dataset must be used for train.py and eval.py. Here, we used --mechanism news to generate the dataset. Other options are --mechanism moore and --mechanism diffdrive.

III. Train a model

You can train a VIN with iteration count K=15 and kernel size F=5 on the datafile mazes.npz by running:

python train.py --datafile mazes.npz --mechanism news --model models.VIN --k 15 --f 5 --save-directory log/vin-k15-f5

This will save outputs to the subdirectory vin-k15-f5/, including the trained models and learning plots.

Similarly, you can train a GPPN by running:

python train.py --datafile mazes.npz --mechanism news --model models.GPPN --k 15 --f 5 --save-directory log/gppn-k15-f5

Notes:

--mechanism must be the same as the one used to generate mazes.npz (which is news in this example).
--f must be an odd integer.

IV. Evaluate a trained model

Once you have a trained VIN model, you can evaluate it on a dataset by running:

python eval.py --datafile mazes.npz --mechanism news --model models.VIN --k 15 --f 5 --load-file log/vin-k15-f5/planner.final.pth

Similarly for GPPN:

python eval.py --datafile mazes.npz --mechanism news --model models.GPPN --k 15 --f 5 --load-file log/gppn-k15-f5/planner.final.pth

Notes:

--mechanism must be the same as the one used to generate mazes.npz (which is news in this example).
--f must be the same as the one used to train the model.

Replicating experiments from our paper

I. Download 2D maze datasets

To replicate experiments from our ICML 2018 paper, first download the datasets by running:

./download_datasets.sh

This will create a subdirectory mazes/ containing the following 2D maze datasets used in our experiments:

datafile	maze size	mechanism	train size	val size	test size
`m15_news_10k.npz`	`15x15`	`news`	10000	2000	2000
`m15_news_25k.npz`	`15x15`	`news`	25000	5000	5000
`m15_news_100k.npz`	`15x15`	`news`	100000	10000	10000
`m15_moore_10k.npz`	`15x15`	`moore`	10000	2000	2000
`m15_moore_25k.npz`	`15x15`	`moore`	25000	5000	5000
`m15_moore_100k.npz`	`15x15`	`moore`	100000	10000	10000
`m15_diffdrive_10k.npz`	`15x15`	`diffdrive`	10000	2000	2000
`m15_diffdrive_25k.npz`	`15x15`	`diffdrive`	25000	5000	5000
`m15_diffdrive_100k.npz`	`15x15`	`diffdrive`	100000	10000	10000
`m28_news_25k.npz`	`28x28`	`news`	25000	5000	5000
`m28_moore_25k.npz`	`28x28`	`moore`	25000	5000	5000
`m28_diffdrive_25k.npz`	`28x28`	`diffdrive`	25000	5000	5000

II. Train VIN and GPPN

Then you can train VIN with the best (K, F) settings for each dataset from our paper by running:

python train.py --datafile mazes/m15_news_10k.npz --mechanism news --model models.VIN --k 30 --f 5
python train.py --datafile mazes/m15_news_25k.npz --mechanism news --model models.VIN --k 20 --f 5
python train.py --datafile mazes/m15_news_100k.npz --mechanism news --model models.VIN --k 30 --f 3

python train.py --datafile mazes/m15_moore_10k.npz --mechanism moore --model models.VIN --k 30 --f 11
python train.py --datafile mazes/m15_moore_25k.npz --mechanism moore --model models.VIN --k 30 --f 5
python train.py --datafile mazes/m15_moore_100k.npz --mechanism moore --model models.VIN --k 30 --f 5

python train.py --datafile mazes/m15_diffdrive_10k.npz --mechanism diffdrive --model models.VIN --k 30 --f 3
python train.py --datafile mazes/m15_diffdrive_25k.npz --mechanism diffdrive --model models.VIN --k 30 --f 3
python train.py --datafile mazes/m15_diffdrive_100k.npz --mechanism diffdrive --model models.VIN --k 30 --f 3

python train.py --datafile mazes/m28_news_25k.npz --mechanism news --model models.VIN --k 56 --f 3
python train.py --datafile mazes/m28_moore_25k.npz --mechanism moore --model models.VIN --k 56 --f 5
python train.py --datafile mazes/m28_diffdrive_25k.npz --mechanism diffdrive --model models.VIN --k 56 --f 3

Similarly, you can train GPPN with the best (K, F) settings for each dataset from our paper by running:

python train.py --datafile mazes/m15_news_10k.npz --mechanism news --model models.GPPN --k 20 --f 9
python train.py --datafile mazes/m15_news_25k.npz --mechanism news --model models.GPPN --k 20 --f 11
python train.py --datafile mazes/m15_news_100k.npz --mechanism news --model models.GPPN --k 30 --f 11

python train.py --datafile mazes/m15_moore_10k.npz --mechanism moore --model models.GPPN --k 30 --f 7
python train.py --datafile mazes/m15_moore_25k.npz --mechanism moore --model models.GPPN --k 30 --f 9
python train.py --datafile mazes/m15_moore_100k.npz --mechanism moore --model models.GPPN --k 30 --f 7

python train.py --datafile mazes/m15_diffdrive_10k.npz --mechanism diffdrive --model models.GPPN --k 30 --f 11
python train.py --datafile mazes/m15_diffdrive_25k.npz --mechanism diffdrive --model models.GPPN --k 30 --f 9
python train.py --datafile mazes/m15_diffdrive_100k.npz --mechanism diffdrive --model models.GPPN --k 30 --f 9

python train.py --datafile mazes/m28_news_25k.npz --mechanism news --model models.GPPN --k 56 --f 11
python train.py --datafile mazes/m28_moore_25k.npz --mechanism moore --model models.GPPN --k 56 --f 9
python train.py --datafile mazes/m28_diffdrive_25k.npz --mechanism diffdrive --model models.GPPN --k 56 --f 11

III. Test Performance Results

Here are the test performance results from running the above commands inside the Docker container lileee/ubuntu-16.04-pytorch-0.4.0-gpu:v1:

	VIN				GPPN
datafile	K	F	%Opt	%Suc	K	F	%Opt	%Suc
`m15_news_10k.npz`	30	5	77.4	79.0	20	9	96.8	97.8
`m15_news_25k.npz`	20	5	83.6	84.2	20	11	99.0	99.3
`m15_news_100k.npz`	30	3	92.6	92.8	30	11	99.7	99.8
`m15_moore_10k.npz`	30	11	86.0	89.3	30	7	97.0	98.0
`m15_moore_25k.npz`	30	5	85.4	88.1	30	9	98.9	99.5
`m15_moore_100k.npz`	30	5	96.9	97.5	30	7	99.6	99.8
`m15_diffdrive_10k.npz`	30	3	98.4	99.0	30	11	99.1	99.7
`m15_diffdrive_25k.npz`	30	3	96.1	98.5	30	9	98.9	99.5
`m15_diffdrive_100k.npz`	30	3	99.0	99.4	30	9	99.8	99.9
`m28_news_25k.npz`	56	3	83.4	84.2	56	11	96.5	97.8
`m28_moore_25k.npz`	56	5	73.3	81.0	56	9	96.5	97.9
`m28_diffdrive_25k.npz`	56	3	82.0	93.6	56	11	95.3	98.0

Feel free to play around with different iteration counts --k and kernel sizes --f.

IV. Version differences (Python, PyTorch)

The test performance results above are slightly different from what is reported in our ICML 2018 paper due to version differences in Python (3.6.5 vs. 2.7.12) and PyTorch (0.4.0 vs. 0.3.1).

Below, we provide instructions to exactly replicate the numbers reported in our ICML 2018 paper.

Checkout the Git branch icml2018:

git checkout icml2018

Load the Docker container used in our experiments by running:

# PyTorch with GPU support
nvidia-docker pull lileee/python-2.7-pytorch-0.3.1-custom:latest
nvidia-docker run -v $(pwd):/home --rm -ti lileee/python-2.7-pytorch-0.3.1-custom:latest

This Docker container uses Python 2.7.12 and a custom version of PyTorch 0.3.1 compiled from source at https://github.com/eparisotto/pytorch.

Train a model:

python train.py --datafile mazes/m15_news_25k.npz --mechanism news --model models.VIN --k 20 --f 5

Citation

If you found this code useful in your research, please cite:

@inproceedings{gppn2018,
  author    = {Lisa Lee and Emilio Parisotto and Devendra Singh Chaplot and Eric Xing and Ruslan Salakhutdinov},
  title     = {Gated Path Planning Networks},
  booktitle = {Proceedings of the 35th International Conference on Machine Learning (ICML 2018)},
  year      = {2018}
}

Acknowledgments

Thanks to @kentsommer for releasing a PyTorch implementation of the original VIN results, which served as a starting point for this codebase.

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
models		models
utils		utils
.gitignore		.gitignore
README.md		README.md
download_datasets.sh		download_datasets.sh
eval.py		eval.py
generate_dataset.py		generate_dataset.py
requirements.txt		requirements.txt
train.py		train.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Gated Path Planning Networks (ICML 2018)

Getting Started

I. Docker container

II. Generate a 2D maze dataset

III. Train a model

IV. Evaluate a trained model

Replicating experiments from our paper

I. Download 2D maze datasets

II. Train VIN and GPPN

III. Test Performance Results

IV. Version differences (Python, PyTorch)

Citation

Acknowledgments

About

Releases

Packages

Contributors 2

Languages

RLAgent/gated-path-planning-networks

Folders and files

Latest commit

History

Repository files navigation

Gated Path Planning Networks (ICML 2018)

Getting Started

I. Docker container

II. Generate a 2D maze dataset

III. Train a model

IV. Evaluate a trained model

Replicating experiments from our paper

I. Download 2D maze datasets

II. Train VIN and GPPN

III. Test Performance Results

IV. Version differences (Python, PyTorch)

Citation

Acknowledgments

About

Resources

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Languages

Packages