Train an RL agent to execute natural language instructions in a 3D Environment (PyTorch)

Last update: Nov 05, 2022

Overview

Gated-Attention Architectures for Task-Oriented Language Grounding

This is a PyTorch implementation of the AAAI-18 paper:

Gated-Attention Architectures for Task-Oriented Language Grounding
Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Rajagopal, Ruslan Salakhutdinov
Carnegie Mellon University

Project Website: https://sites.google.com/view/gated-attention

This repository contains:

Code for training an A3C-LSTM agent using Gated-Attention
Code for Doom-based language grounding environment

Dependencies

ViZDoom
PyTorch
Opencv

(We recommend using Anaconda)

Usage

Using the Environment

For running a random agent:

python env_test.py

To play in the environment:

python env_test.py --interactive 1

To change the difficulty of the environment (easy/medium/hard):

python env_test.py -d easy

Training Gated-Attention A3C-LSTM agent

For training a A3C-LSTM agent with 32 threads:

python a3c_main.py --num-processes 32 --evaluate 0

The code will save the best model at ./saved/model_best.

To the test the pre-trained model for Multitask Generalization:

python a3c_main.py --evaluate 1 --load saved/pretrained_model

To the test the pre-trained model for Zero-shot Task Generalization:

python a3c_main.py --evaluate 2 --load saved/pretrained_model

To the visualize the model while testing add '--visualize 1':

python a3c_main.py --evaluate 2 --load saved/pretrained_model --visualize 1

To test the trained model, use --load saved/model_best in the above commands.

All arguments for a3c_main.py:

  -h, --help            show this help message and exit
  -l MAX_EPISODE_LENGTH, --max-episode-length MAX_EPISODE_LENGTH
                        maximum length of an episode (default: 30)
  -d DIFFICULTY, --difficulty DIFFICULTY
                        Difficulty of the environment, "easy", "medium" or
                        "hard" (default: hard)
  --living-reward LIVING_REWARD
                        Default reward at each time step (default: 0, change
                        to -0.005 to encourage shorter paths)
  --frame-width FRAME_WIDTH
                        Frame width (default: 300)
  --frame-height FRAME_HEIGHT
                        Frame height (default: 168)
  -v VISUALIZE, --visualize VISUALIZE
                        Visualize the envrionment (default: 0, use 0 for
                        faster training)
  --sleep SLEEP         Sleep between frames for better visualization
                        (default: 0)
  --scenario-path SCENARIO_PATH
                        Doom scenario file to load (default: maps/room.wad)
  --interactive INTERACTIVE
                        Interactive mode enables human to play (default: 0)
  --all-instr-file ALL_INSTR_FILE
                        All instructions file (default:
                        data/instructions_all.json)
  --train-instr-file TRAIN_INSTR_FILE
                        Train instructions file (default:
                        data/instructions_train.json)
  --test-instr-file TEST_INSTR_FILE
                        Test instructions file (default:
                        data/instructions_test.json)
  --object-size-file OBJECT_SIZE_FILE
                        Object size file (default: data/object_sizes.txt)
  --lr LR               learning rate (default: 0.001)
  --gamma G             discount factor for rewards (default: 0.99)
  --tau T               parameter for GAE (default: 1.00)
  --seed S              random seed (default: 1)
  -n N, --num-processes N
                        how many training processes to use (default: 4)
  --num-steps NS        number of forward steps in A3C (default: 20)
  --load LOAD           model path to load, 0 to not reload (default: 0)
  -e EVALUATE, --evaluate EVALUATE
                        0:Train, 1:Evaluate MultiTask Generalization
                        2:Evaluate Zero-shot Generalization (default: 0)
  --dump-location DUMP_LOCATION
                        path to dump models and log (default: ./saved/)

Demostration videos:

Multitask Generalization video: https://www.youtube.com/watch?v=YJG8fwkv7gA

Zero-shot Task Generalization video: https://www.youtube.com/watch?v=JziCKsLrudE

Different stages of training: https://www.youtube.com/watch?v=o_G6was03N0

Cite as

Chaplot, D.S., Sathyendra, K.M., Pasumarthi, R.K., Rajagopal, D. and Salakhutdinov, R., 2017. Gated-Attention Architectures for Task-Oriented Language Grounding. arXiv preprint arXiv:1706.07230. (PDF)

Bibtex:

@article{chaplot2017gated,
  title={Gated-Attention Architectures for Task-Oriented Language Grounding},
  author={Chaplot, Devendra Singh and Sathyendra, Kanthashree Mysore and Pasumarthi, Rama Kumar and Rajagopal, Dheeraj and Salakhutdinov, Ruslan},
  journal={arXiv preprint arXiv:1706.07230},
  year={2017}
}

Acknowledgements

This repository uses ViZDoom API (https://github.com/mwydmuch/ViZDoom) and parts of the code from the API. The implementation of A3C is borrowed from https://github.com/ikostrikov/pytorch-a3c. The poisson-disc code is borrowed from https://github.com/IHautaI/poisson-disc.

Train an RL agent to execute natural language instructions in a 3D Environment (PyTorch)

Related tags

Overview

Gated-Attention Architectures for Task-Oriented Language Grounding

This repository contains:

Dependencies

Usage

Using the Environment

Training Gated-Attention A3C-LSTM agent

Demostration videos:

Cite as

Bibtex:

Acknowledgements

Owner

Devendra Chaplot

Implementation of Graph Convolutional Networks in TensorFlow

Java and SHACL code commented in the paper "Towards compliance checking in reified I/O logic via SHACL" submitted to ICAIL 2021

Graph WaveNet apdapted for brain connectivity analysis.

A platform to display the carbon neutralization information for researchers, decision-makers, and other participants in the community.

Multi Task Vision and Language

TensorFlow implementation of Adaptive Information Transfer Multi-task (AITM) framework. Code for the paper submitted to KDD21: Modeling the Sequential Dependence among Audience Multi-step Conversions with Multi-task Learning for Customer Acquisition.

GraphGT: Machine Learning Datasets for Graph Generation and Transformation

CRISCE: Automatically Generating Critical Driving Scenarios From Car Accident Sketches

Official PyTorch Implementation of GAN-Supervised Dense Visual Alignment

A tool to analyze leveraged liquidity mining and find optimal option combination for hedging.

Optimizing Value-at-Risk and Conditional Value-at-Risk of Black Box Functions with Lacing Values (LV)

Repository relating to the CVPR21 paper TimeLens: Event-based Video Frame Interpolation

torchbearer: A model fitting library for PyTorch

Learning an Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems

paper: Hyperspectral Remote Sensing Image Classification Using Deep Convolutional Capsule Network

PyTorch Implementation of Realtime Multi-Person Pose Estimation project.

Implementation of Invariant Point Attention, used for coordinate refinement in the structure module of Alphafold2, as a standalone Pytorch module

Prototype-based Incremental Few-Shot Semantic Segmentation

The official implementation of Variable-Length Piano Infilling (VLI).

Lightweight mmm - Lightweight (Bayesian) Media Mix Model