mapnet-pytorch

This is an unofficial PyTorch re-implementation of MapNet: An Allocentric Spatial Memory for Mapping Environments (CVPR 2018, oral). It reproduces some preliminary results on Maze and AVD environments. Some parts of the code were borrowed from the original implementation provided by the author at https://github.com/jotaf98/mapnet.

Requirements

The following are the primary requirements:

• PyTorch 1.2
• Python 3.6
• PyTorch scatter (https://github.com/rusty1s/pytorch_scatter)
• Einops (https://github.com/arogozhnikov/einops)
• TensorboardX

Datasets

• Maze dataset: https://github.com/jotaf98/mapnet/tree/master/data/maze
• Active Vision Dataset (processed)

Setup

Add the following to ~/.bashrc

export MAPNET_PATH=<PATH TO mapnet-pytorch>
export PYTHONPATH=$PYTHONPATH:$MAPNET_PATH

Usage

Training on Maze environment

cd $MAPNET_PATH
mkdir -p trained_models/maze_models

python train.py --lr 3e-3 --seed 1 --batch-size 100 --num-steps 5 \
                --save-interval 1000 --eval-interval 1000 --num-updates 100000 \
                --env-name maze --data-path <PATH TO mazes-10-10-100000.txt> \
                --save-dir trained_models/maze_models --log-dir trained_models/maze_models/ \ 
                --map-size 15 --local-map-size 11 --map-scale 1.0 --log-interval 100

Training on Active Vision Dataset

cd $MAPNET_PATH
mkdir -p trained_models/avd_models

python train.py --lr 1e-3 --seed 1 --batch-size 64 --num-steps 5 \
                --save-interval 1000 --eval-interval 1000 --num-updates 100000 \
                --env-name avd --data-path <PATH TO avd_trajectories_T20.h5> \
                --save-dir trained_models/avd_models --log-dir trained_models/avd_models \
                --map-size 31 --local-map-size 21 --map-scale 300.0 --log-interval 100