Learning from Guided Play: A Scheduled Hierarchical Approach for Improving Exploration in Adversarial Imitation Learning Source Code

Related tags

Deep Learninglfgp
Overview

Learning from Guided Play: A Scheduled Hierarchical Approach for Improving Exploration in Adversarial Imitation Learning

Trevor Ablett*, Bryan Chan*, Jonathan Kelly (*equal contribution)

Poster at Neurips 2021 Deep Reinforcement Learning Workshop


Adversarial Imitation Learning (AIL) is a technique for learning from demonstrations that helps remedy the distribution shift problem that occurs with Behavioural Cloning. Empirically, we found that for manipulation tasks, off-policy AIL can suffer from inefficient or stagnated learning. In this work, we resolve this by enforcing exploration of a set of easy-to-define auxiliary tasks, in addition to a main task.

This repository contains the source code for reproducing our results.

Setup

We recommend the readers set up a virtual environment (e.g. virtualenv, conda, pyenv, etc.). Please also ensure to use Python 3.7 as we have not tested in any other Python versions. In the following, we assume the working directory is the directory containing this README:

.
├── lfgp_data/
├── liegroups/
├── manipulator-learning/
├── rl_sandbox/
├── README.md
└── requirements.txt

To install, simply clone and install with pip, which will automatically install all dependencies:

git clone [email protected]:utiasSTARS/lfgp.git && cd lfgp
pip install rl_sandbox

Environments

In this paper, we evaluated our method in the four environments listed below:

bring_0                  # bring blue block to blue zone
stack_0                  # stack blue block onto green block
insert_0                 # insert blue block into blue zone slot
unstack_stack_env_only_0 # remove green block from blue block, and stack blue block onto green block

Trained Models and Expert Data

The expert and trained lfgp models can be found at this google drive link. The zip file is 570MB. All of our generated expert data is included, but we only include single seeds of each trained model to reduce the size.

The Data Directory

This subsection provides the desired directory structure that we will be assuming for the remaining README. The unzipped lfgp_data directory follows the structure:

.
├── lfgp_data/
│   ├── expert_data/
│   │   ├── unstack_stack_env_only_0-expert_data/
│   │   │   ├── reset/
│   │   │   │   ├── 54000_steps/
│   │   │   │   └── 9000_steps/
│   │   │   └── play/
│   │   │       └── 9000_steps/
│   │   ├── stack_0-expert_data/
│   │   │   └── (same as unstack_stack_env_only_0-expert_data)/
│   │   ├── insert_0-expert_data/
│   │   │   └── (same as unstack_stack_env_only_0-expert_data)/
│   │   └── bring_0-expert_data/
│   │       └── (same as unstack_stack_env_only_0-expert_data)/
│   └── trained_models/
│       ├── experts/
│       │   ├── unstack_stack_env_only_0/
│       │   ├── stack_0/
│       │   ├── insert_0/
│       │   └── bring_0/
│       ├── unstack_stack_env_only_0/
│       │   ├── multitask_bc/
│       │   ├── lfgp_ns/
│       │   ├── lfgp/
│       │   ├── dac/
│       │   ├── bc_less_data/
│       │   └── bc/
│       ├── stack_0/
│       │   └── (same as unstack_stack_env_only_0)
│       ├── insert_0/
│       │   └── (same as unstack_stack_env_only_0)
│       └── bring_0/
│           └── (same as unstack_stack_env_only_0)
├── liegroups/
├── manipulator-learning/
├── rl_sandbox/
├── README.md
└── requirements.txt

Create Expert and Generate Expert Demonstrations

Readers can generate their own experts and expert demonstrations by executing the scripts in the rl_sandbox/rl_sandbox/examples/lfgp/experts directory. More specifically, create_expert.py and create_expert_data.py respectively train the expert and generate the expert demonstrations. We note that training the expert is time consuming and may take up to multiple days.

To create an expert, you can run the following command:

# Create a stack expert using SAC-X with seed 0. --gpu_buffer would store the replay buffer on the GPU.
# For more details, please use --help command for more options.
python rl_sandbox/rl_sandbox/examples/lfgp/experts/create_expert.py \
    --seed=0 \
    --main_task=stack_0 \
    --device=cuda \
    --gpu_buffer

A results directory will be generated. A tensorboard, an experiment setting, a training progress file, model checkpoints, and a buffer checkpoint will be created.

To generate play-based and reset-based expert data using a trained model, you can run the following commands:

# Generate play-based stack expert data with seed 1. The program halts when one of --num_episodes or --num_steps is satisfied.
# For more details, please use --help command for more options
python rl_sandbox/rl_sandbox/examples/lfgp/experts/create_expert_data.py \
--model_path=data/stack_0/expert/state_dict.pt \
--config_path=data/stack_0/expert/sacx_experiment_setting.pkl \
--save_path=./test_expert_data \
--num_episodes=10 \
--num_steps=1000 \
--seed=1 \
--render

# Generate reset-based stack expert data with seed 1. Note that --num_episodes will need to be scaled by number of tasks (i.e. num_episodes * num_tasks).
python rl_sandbox/rl_sandbox/examples/lfgp/experts/create_expert_data.py \
--model_path=data/stack_0/expert/state_dict.pt \
--config_path=data/stack_0/expert/sacx_experiment_setting.pkl \
--save_path=./test_expert_data \
--num_episodes=10 \
--num_steps=1000 \
--seed=1 \
--render \
--reset_between_intentions

The generated expert data will be stored under --save_path, in separate files int_0.gz, ..., int_{num_tasks - 1}.gz.

Training the Models with Imitation Learning

In the following, we assume the expert data is generated following the previous section and is stored under test_expert_data. The training scripts run_*.py are stored in rl_sandbox/rl_sandbox/examples/lfgp directory. There are five run scripts, each corresponding to a variant of the compared methods (except for behavioural cloning less data, since the change is only in the expert data). The runs will be saved in the same results directory mentioned previously. Note that the default hyperparameters specified in the scripts are listed on the appendix.

Behavioural Cloning (BC)

There are two scripts for single-task and multitask BC: run_bc.py and run_multitask_bc.py. You can run the following commands:

# Train single-task BC agent to stack with using reset-based data.
# NOTE: intention 2 is the main intention (i.e. stack intention). The main intention is indexed at 2 for all environments.
python rl_sandbox/rl_sandbox/examples/lfgp/run_bc.py \
--seed=0 \
--expert_path=test_expert_data/int_2.gz \
--main_task=stack_0 \
--render \
--device=cuda

# Train multitask BC agent to stack with using reset-based data.
python rl_sandbox/rl_sandbox/examples/lfgp/run_multitask_bc.py \
--seed=0 \
--expert_paths=test_expert_data/int_0.gz,\
test_expert_data/int_1.gz,\
test_expert_data/int_2.gz,\
test_expert_data/int_3.gz,\
test_expert_data/int_4.gz,\
test_expert_data/int_5.gz
--main_task=stack_0 \
--render \
--device=cuda

Adversarial Imitation learning (AIL)

There are three scripts for Discriminator-Actor-Critic (DAC), Learning from Guided Play (LfGP), and LfGP-NS (No Schedule): run_dac.py, run_lfgp.py, run_lfgp_ns.py. You can run the following commands:

# Train DAC agent to stack with using reset-based data.
python rl_sandbox/rl_sandbox/examples/lfgp/run_dac.py \
--seed=0 \
--expert_path=test_expert_data/int_2.gz \
--main_task=stack_0 \
--render \
--device=cuda

# Train LfGP agent to stack with using reset-based data.
python rl_sandbox/rl_sandbox/examples/lfgp/run_lfgp.py \
--seed=0 \
--expert_paths=test_expert_data/int_0.gz,\
test_expert_data/int_1.gz,\
test_expert_data/int_2.gz,\
test_expert_data/int_3.gz,\
test_expert_data/int_4.gz,\
test_expert_data/int_5.gz
--main_task=stack_0 \
--device=cuda \
--render

# Train LfGP-NS agent to stack with using reset-based data.
python rl_sandbox/rl_sandbox/examples/lfgp/run_lfgp_ns.py \
--seed=0 \
--expert_paths=test_expert_data/int_0.gz,\
test_expert_data/int_1.gz,\
test_expert_data/int_2.gz,\
test_expert_data/int_3.gz,\
test_expert_data/int_4.gz,\
test_expert_data/int_5.gz,\
test_expert_data/int_6.gz \
--main_task=stack_0 \
--device=cuda \
--render

Evaluating the Models

The readers may load up trained agents and evaluate them using the evaluate.py script under the rl_sandbox/rl_sandbox/examples/eval_tools directory. Currently, only the lfgp agent is supplied due to the space restrictions mentioned above.

# For single-task agents - DAC, BC
# To run single-task agent (e.g. BC)
python rl_sandbox/rl_sandbox/examples/eval_tools/evaluate.py \
--seed=1 \
--model_path=data/stack_0/il_agents/bc/state_dict.pt \
--config_path=data/stack_0/il_agents/bc/bc_experiment_setting.pkl \
--num_episodes=5 \
--intention=0 \
--render \
--device=cuda

# For multitask agents - SAC-X, LfGP, LfGP-NS, Multitask BC
# To run all intentions for multitask agents (e.g. SAC-X)
python rl_sandbox/rl_sandbox/examples/eval_tools/evaluate.py \
--seed=1 \
--model_path=data/stack_0/expert/state_dict.pt \
--config_path=data/stack_0/expert/sacx_experiment_setting.pkl \
--num_episodes=5 \
--intention=-1 \
--render \
--device=cuda

# To run only the main intention for multitask agents (e.g. LfGP)
python rl_sandbox/rl_sandbox/examples/eval_tools/evaluate.py \
--seed=1 \
--model_path=data/stack_0/il_agents/lfgp/state_dict.pt \
--config_path=data/stack_0/il_agents/lfgp/lfgp_experiment_setting.pkl \
--num_episodes=5 \
--intention=2 \
--render \
--device=cuda

Owner
STARS Laboratory
We are the Space and Terrestrial Autonomous Robotic Systems Laboratory at the University of Toronto
STARS Laboratory
Molecular AutoEncoder in PyTorch

MolEncoder Molecular AutoEncoder in PyTorch Install $ git clone https://github.com/cxhernandez/molencoder.git && cd molencoder $ python setup.py insta

Carlos Hernández 80 Dec 05, 2022
Noise Conditional Score Networks (NeurIPS 2019, Oral)

Generative Modeling by Estimating Gradients of the Data Distribution This repo contains the official implementation for the NeurIPS 2019 paper Generat

451 Dec 26, 2022
A style-based Quantum Generative Adversarial Network

Style-qGAN A style based Quantum Generative Adversarial Network (style-qGAN) model for Monte Carlo event generation. Tutorial We have prepared a noteb

9 Nov 24, 2022
Implementation of the Chamfer Distance as a module for pyTorch

Chamfer Distance for pyTorch This is an implementation of the Chamfer Distance as a module for pyTorch. It is written as a custom C++/CUDA extension.

Christian Diller 205 Jan 05, 2023
All course materials for the Zero to Mastery Deep Learning with TensorFlow course.

All course materials for the Zero to Mastery Deep Learning with TensorFlow course.

Daniel Bourke 3.4k Jan 07, 2023
VISSL is FAIR's library of extensible, modular and scalable components for SOTA Self-Supervised Learning with images.

What's New Below we share, in reverse chronological order, the updates and new releases in VISSL. All VISSL releases are available here. [Oct 2021]: V

Meta Research 2.9k Jan 07, 2023
This is the code for "HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields".

HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields This is the code for "HyperNeRF: A Higher-Dimensional

Google 702 Jan 02, 2023
An implementation of the AlphaZero algorithm for Gomoku (also called Gobang or Five in a Row)

AlphaZero-Gomoku This is an implementation of the AlphaZero algorithm for playing the simple board game Gomoku (also called Gobang or Five in a Row) f

Junxiao Song 2.8k Dec 26, 2022
TransCD: Scene Change Detection via Transformer-based Architecture

TransCD: Scene Change Detection via Transformer-based Architecture

wangzhixue 29 Dec 11, 2022
S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided Distribution Calibration (CVPR 2021)

S2-BNN (Self-supervised Binary Neural Networks Using Distillation Loss) This is the official pytorch implementation of our paper: "S2-BNN: Bridging th

Zhiqiang Shen 52 Dec 24, 2022
some academic posters as references. May we have in-person poster session soon!

some academic posters as references. May we have in-person poster session soon!

Bolei Zhou 472 Jan 06, 2023
This is an official implementation for "SimMIM: A Simple Framework for Masked Image Modeling".

SimMIM By Zhenda Xie*, Zheng Zhang*, Yue Cao*, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai and Han Hu*. This repo is the official implementation of

Microsoft 674 Dec 26, 2022
This is the reference implementation for "Coresets via Bilevel Optimization for Continual Learning and Streaming"

Coresets via Bilevel Optimization This is the reference implementation for "Coresets via Bilevel Optimization for Continual Learning and Streaming" ht

Zalán Borsos 51 Dec 30, 2022
2D Human Pose estimation using transformers. Implementation in Pytorch

PE-former: Pose Estimation Transformer Vision transformer architectures perform very well for image classification tasks. Efforts to solve more challe

Panteleris Paschalis 23 Oct 17, 2022
Barlow Twins and HSIC

Barlow Twins and HSIC Unofficial Pytorch implementation for Barlow Twins and HSIC_SSL on small datasets (CIFAR10, STL10, and Tiny ImageNet). Correspon

Yao-Hung Hubert Tsai 49 Nov 24, 2022
Markov Attention Models

Introduction This repo contains code for reproducing the results in the paper Graphical Models with Attention for Context-Specific Independence and an

Vicarious 0 Dec 09, 2021
Graph Attention Networks

GAT Graph Attention Networks (Veličković et al., ICLR 2018): https://arxiv.org/abs/1710.10903 GAT layer t-SNE + Attention coefficients on Cora Overvie

Petar Veličković 2.6k Jan 05, 2023
Video Frame Interpolation without Temporal Priors (a general method for blurry video interpolation)

Video Frame Interpolation without Temporal Priors (NeurIPS2020) [Paper] [video] How to run Prerequisites NVIDIA GPU + CUDA 9.0 + CuDNN 7.6.5 Pytorch 1

YoujianZhang 31 Sep 04, 2022
Official code repository for the EMNLP 2021 paper

Integrating Visuospatial, Linguistic and Commonsense Structure into Story Visualization PyTorch code for the EMNLP 2021 paper "Integrating Visuospatia

Adyasha Maharana 23 Dec 19, 2022
This is the repository for the NeurIPS-21 paper [Contrastive Graph Poisson Networks: Semi-Supervised Learning with Extremely Limited Labels].

CGPN This is the repository for the NeurIPS-21 paper [Contrastive Graph Poisson Networks: Semi-Supervised Learning with Extremely Limited Labels]. Req

10 Sep 12, 2022