Framework for training options with different attention mechanism and using them to solve downstream tasks.

Last update: Nov 03, 2022

Related tags

Deep Learning hrl_attention

Overview

Using Attention in HRL

Framework for training options with different attention mechanism and using them to solve downstream tasks.

Requirements

GPU required

conda env create -f conda_env.yml

After the instalation ends you can activate your environment and install remaining dependencies. (e.g. sub-module gym_minigrid which is a modified version of MiniGrid )

conda activate affenv
cd gym-minigrid
pip install -e .
cd ../
pip install -e .

Instructions

In order to train options and IC_net follow these steps:

1. Configure desired environment - number of task and objects per task in file config/op_ic_net.yaml. E.g:
  env_args:
    task_size: 3
    num_tasks: 4

2. Configure desired type of attention (between "affordance", "interest", "nan") - in file config/op_ic_net.yaml. E.g. 
main:
  attention: "affordance" 

3. Train by running command
liftoff train_main.py configs/op_ic_net.yaml

Once a pre-trained option checkpoint exists a HRL agent can be trained to solve the downstream task (for the same environment the options were trained on). Follow these steps in order to train an HRL-Agent with different types of attentions:

1. Configure checkpoint (experiment config file and options_model_id) for pre-trained Options and IC_net - in file configs/hrl-agent.yaml. E.g: 

main:
  options_model_cfg: "results/op_aff_4x3/0000_multiobj/0/cfg.yaml"
  options_model_id: -1  # Last checkpoint will be used

2. Configure type of attention for training the HRL-agent (between "affordance", "interest", "nan") - in file configs/hrl-agent.yaml. E.g:
main:
  modulate_policy: affordance

3. Train HRL-agent by running command
liftoff train_mtop_ppo.py configs/hrl-agent.yaml

Both training scrips produce results in the results folder, where all the outputs are going to be stored including train/eval logs, checkpoints. Live plotting is integrated using services from Wandb (plotting has to be enabled in the config file main:plot and user logged in Wandb or user login api key in the file .wandb_key).

The console output is also available in a form:

Option Pre-training e.g.:

U 11 | F 022528 | FPS 0024 | D 402 | rR:u, 0.03 | F:u, 41.77 | tL:u 0.00 | tPL:u 6.47 | tNL:u 0.00 | t 52 | aff_loss 0.0570 | aff 2.8628 | NOaff 0.0159 | ic 0.0312 | cnt_ic 1.0000 | oe 2.4464 | oic0 0.0000 | oic1 0.0000 | oic2 0.0000 | oic3 0.0000 | oPic0 0.0000 | oPic1 0.0000 | oPic2 0.0000 | oPic3 0.0000 | icB 0.0208 | PicB 0.1429 | icND 0.0192

Some of the training entries decodes as

F - number of frames (steps in the env)
tL - termination loss
aff_loss - IC_net loss
cnt_ic - Intent completion per training batch 
oicN - Intent completion fraction for each option N out of Total option N sampled
oPicN - Intent completion fraction for each option N out of affordable ones
PicB - Intent completion average over all options out of affordable ones

HRL-agent training

U 1 | F 4555192.0 | FPS 21767 | D 209 | rR:u, 0.00 | F:u, 8.11 | e:u, 2.48 | v:u 0.00 | pL:u 0.01 | vL:u 0.00 | g:u 0.01 | TrR:u, 0.00

Some of the training entries decodes as

F - number of frames (steps in the env offseted by the number of pre-training steps)
rR - Accumulated episode reward average
TrR - Average episode success rate

Framework structure

The code is organised as follows:

agents/ - implementation of agents (e.g. training options and IC_net multistep_affordance.py; hrl-agent PPO ppo_smdp.py )
configs/ - config files for training agents
gym-minigrid/ - sub-module - Minigrid envs
models/ - Neural network modules (e.g options with IC_net aff_multistep.py and CNN backbone extractor_cnn_v2.py)
utils/ - Scripts for e.g.: running envs in parallel, preprocessing observations, gym wrappers, data structures, logging modules
train_main.py - Train Options with IC_net
train_mtop_ppo.py - Train HRL-agent

Acknowledgements

We used PyTorch as a machine learning framework.

We used liftoff for experiment management.

We used wandb for plotting.

We used PPO adapted for training our agents.

We used MiniGrid to create our environment.

Framework for training options with different attention mechanism and using them to solve downstream tasks.

Related tags

Overview

Using Attention in HRL

Requirements

Instructions

Framework structure

Acknowledgements

Owner

paper list in the area of reinforcenment learning for recommendation systems

The source codes for ACL 2021 paper 'BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data'

Reference models and tools for Cloud TPUs.

CBKH: The Cornell Biomedical Knowledge Hub

Numba-accelerated Pythonic implementation of MPDATA with examples in Python, Julia and Matlab

Deep Reinforcement Learning for mobile robot navigation in ROS Gazebo simulator

Performance Analysis of Multi-user NOMA Wireless-Powered mMTC Networks: A Stochastic Geometry Approach

Deep generative modeling for time-stamped heterogeneous data, enabling high-fidelity models for a large variety of spatio-temporal domains.

3rd Place Solution of the Traffic4Cast Core Challenge @ NeurIPS 2021

Applying CLIP to Point Cloud Recognition.

Parameterized Explainer for Graph Neural Network

Dynamic Capacity Networks using Tensorflow

ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation

[ICCV 2021 (oral)] Planar Surface Reconstruction from Sparse Views

Github project for Attention-guided Temporal Coherent Video Object Matting.

Official PaddlePaddle implementation of Paint Transformer

This computer program provides a reference implementation of Lagrangian Monte Carlo in metric induced by the Monge patch

Effect of Deep Transfer and Multi task Learning on Sperm Abnormality Detection

This repository contains all data used for writing a research paper Multiple Object Trackers in OpenCV: A Benchmark, presented in ISIE 2021 conference in Kyoto, Japan.

Implementation of C-RNN-GAN.