Learning Object Relation Graph and Tentative Policy for Visual Navigation (Updating)

Overview

Dependencies

• python 3.6+
• python libraries
  • pytorch 1.4+ & tochvision
  • ai2thor
  • networkx
  • numpy
  • scipy
  • tensorboardX

Install the necessary packages. If you are using pip then simply run pip install -r requirements.txt.

Data

Your ~/Data/ folder should look like this

Data
└── AI2thor_offline_data/
    ├── FloorPlan1
    │   ├── resnet18_featuremap.hdf5
    │   ├── graph.json
    │   ├── visible_object_map_1.5.json
    │   ├── det_feature_categories.hdf5
    │   ├── grid.json
    │   └── optimal_action.json
    ├── FloorPlan2
    └── ...

You can download the dataset used in our paper here, test data here and RGB data here.

Training

python main.py --model GraphModel --worker 12 --gpu-ids 0 1 --title a3c_graph

Evaluating

python full_eval.py --model GraphModel --results_json a3c_graph_results.json --gpu-ids 0 1 --title a3c_graph