Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.

README.md

Decentralized Distributed PPO

Provides changes to the core baseline ppo algorithm and training script to implemented Decentralized Distributed PPO (DD-PPO). DD-PPO leverages distributed data parallelism to seamlessly scale PPO to hundreds of GPUs with no centralized server.

See the paper for more detail.

Running

There are two example scripts to run provided. A single node script that leverages torch.distributed.launch to create multiple workers: single_node.sh, and a multi-node script that leverages SLURM to create all the works on multiple nodes: multi_node_slurm.sh.

The two recommended backends are GLOO and NCCL. Use NCCL if your system has it, and GLOO if otherwise.

See pytorch's distributed docs and pytorch's distributed tutorial for more information.

Pretrained Models (PointGoal Navigation with GPS+Compass)

All weights available as a zip here.

Depth models

Architecture Training Data Val SPL Test SPL URL
ResNet50 + LSTM512 Gibson 4+ 0.922 0.917 gibson-4plus-resnet50.pth
ResNet50 + LSTM512 Gibson 4+ and MP3D(train/val/test)
Caution: Trained on MP3D val and test
0.956 0.941 gibson-4plus-mp3d-train-val-test-resnet50.pth
ResNet50 + LSTM512 Gibson 2+ 0.956 0.944 gibson-2plus-resnet50.pth
SE-ResNeXt50 + LSTM512 Gibson 2+ 0.959 0.943 gibson-2plus-se-resneXt101.pth
SE-ResNeXt101 + LSTM1024 Gibson 2+ 0.969 0.948 gibson-2plus-se-resneXt101-lstm1024.pth

RGB models

Architecture Training Data Val SPL Test SPL URL
SE-ResNeXt50 + LSTM512 Gibson 2+ and MP3D(train/val/test)
Caution: Trained on MP3D val and test
0.933 0.920 gibson-2plus-mp3d-train-val-test-se-resneXt50-rgb.pth

Blind Models

Architecture Training Data Val SPL Test SPL URL
LSTM512 Gibson 0+ and MP3D(train/val/test)
Caution: Trained on MP3D val and test
0.729 0.676 gibson-0plus-mp3d-train-val-test-blind.pth

Note: Evaluation was done with sampled actions.

All model weights are subject to Matterport3D Terms-of-Use.

Citing

If you use DD-PPO or the model-weights in your research, please cite the following paper:

@article{wijmans2020ddppo,
  title = {{DD-PPO}: {L}earning Near-Perfect PointGoal Navigators from 2.5 Billion Frames},
  author =  {Erik Wijmans and Abhishek Kadian and Ari Morcos and Stefan Lee and Irfan Essa and Devi Parikh and Manolis Savva and Dhruv Batra},
  journal = {International Conference on Learning Representations (ICLR)},
  year =    {2020}
}
You can’t perform that action at this time.