sx-zhang/gman

GMAN: Generative Meta-Adversarial Network for Unseen Object Navigation (ECCV 2022)

GMAN: Generative Meta-Adversarial Network for Unseen Object Navigation

Setup

• Clone the repository git clone https://github.com/sx-zhang/GMAN.git and move into the top-level directory cd GMAN
• Create conda environment. conda env create -f environment.yml
• Activate the environment. conda activate ng
• We provide pre-trained model of our GMAN in the trained_models directory.
• Download our dataset.
• Then extract files in data by tar -zxvf ai2thor_GMAN.tar.gz
• The data folder should look like this

  data/ 
    └── ai2thor//
        ├── FloorPlan1/
        │   ├── resnet18_featuremap.hdf5
        │   ├── graph.json
        │   ├── visible_object_map.json
        │   ├── att_in_view_v2.hdf5
        │   ├── grid.json
        ├── FloorPlan2/
        └── ...

Training and Evaluation

Train the baseline model

python main.py --title Basemodel --model BaseModel --workers 12 -–gpu-ids 0

Train our GMAN model

python main.py --title GMAN --model GMAN --workers 12 -–gpu-ids 0 --num_steps 20

Evaluate the GMAN model for seen objects

python full_eval.py \
    --title GMAN \
    --model GMAN \
    --results-json GMAN_seen.json \
    --gpu-ids 0 \
    --num_steps 20 \
    --seen seen

Evaluate the GMAN model for unseen objects

python full_eval.py \
    --title GMAN \
    --model GMAN \
    --results-json GMAN_unseen.json \
    --gpu-ids 0 \
    --num_steps 20 \
    --seen unseen

gman's Issues

dataset reference 404 not found

Thank you for providing the dataset! But the link in ReadMe.md seems to be not found.
btw, is the RoboTHOR offline data also in the provided link?

How to achieve the performance of the pre-trained model？

I trained the model according to your code, and the experimental results show a difference of 8% in SR compared to the pre-trained model you provided (best model at the 200,000th iteration). Could you please confirm if there are differences in the hyperparameters between your pre-training model and the ones used in the code? Additionally, could you provide suggestions on how to adjust the hyperparameters to improve performance?

experimental conditions

Would you like to know what your experimental conditions are, how many GPUs, the type of GPU, and how long the training time is, thank you

Dataset download

Thank you so much for open sourcing this great work. But I encountered a problem when downloading the dataset, the corresponding link of the dataset in Readme.md is invalid. How can I get the data? Thank you !

att_in_view_v2.hdf5

Could you plz provide the script about how to generate the dataset, especially the att_in_view_v2.hdf5 file?