Official code for ICCV 2023 paper: GAIT: Generating Aesthetic Indoor Tours with Deep Reinforcement Learning

Installing Dependencies

Since we install habitat-sim by building from source, the environment.yml file we provide could not be directly used but can serve as reference.

1. Start by creating an conda environment with python 3.7, as required by habitat-sim: $ conda create -n habitat python=3.7 cmake=3.14.0
2. Then, install pytorch. If habitat-sim is installed before pytorch, then pytorch would be installed as cpu-only. conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
3. clone habitat-sim github.
4. Edit the habitat-sim source code. By default, habitat-sim has "gravity" enabled, so if we move the agent upwards, it falls down to the ground by itself. We change its source code to disable this.
   1. $ cd habitat-sim/src_python/habitat_sim.
   2. Edit simulator.py. Search for step(, there are three overloaded methods called step(). For each one, add a parameter apply_filter=True. In the third step() definition, go to the line calling agent.act(), and append apply_filter=apply_filter as a parameter.
   3. Edit agent/agent.py. Search for act(, append apply_filter=True as a parameter. In the act() method, there are two calls of self.controls.action(). Append apply_filter=apply_filter to each of them.
5. Install habitat-sim from source. habitat-sim is available as a conda pacakge, but we have met some issues that could be solved by installing from source. Follow this guide from habitat-sim Github.
   1. If you have nvcc: $ export CUDACXX=/usr/local/cuda/bin/nvcc. Otherwise, install nvcc via conda and then export path to it: $sudo apt install libxml2, $ conda install -c conda-forge cudatoolkit-dev, $ export CUDACXX=/home/username/anaconda3/envs/habitat/pkgs/cuda-toolkit/bin/nvcc
   2. sudo apt install build-essential
   3. For $ python setup.py install: use flag --with-cuda, and include --headless flag as needed.
6. While you are building habitat-sim from source, take the time to download the Replica Dataset
   1. $ git clone https://github.com/facebookresearch/Replica-Dataset.git
   2. $ sudo apt install pigz
   3. Make sure Replica-Dataset and gait are in the same folder. Then $cd Replica-Dataset, $./download.sh ../gait/habitat_data/Replica
   4. $ wget http://dl.fbaipublicfiles.com/habitat/sorted_faces.zip, $ unzip sorted_faces.zip
   5. $./sorted_faces/copy_to_folders ~/aestheticview/habitat_data/Replica/
7. Install conda dependencies in environment.yml
8. Install pip dependencies
   1. $ pip install gym tensorboard dm-env termcolor cma
   2. $ pip install -U ray
   3. $ pip install hydra-core --upgrade
   4. $ pip install hydra-submitit-launcher --upgrade
9. Download the pretrained weights of the aesthetics model (View Evaluation Net) from https://github.com/zijunwei/ViewEvaluationNet, and put EvaluationNet.pth.tar in gait/snapshots/params

Training

All hyperparameters are set to the default condition. cfgs/config.yaml contains the hyperparameters for GAIT environment and for DrQv2. Specifically, ray enables or disables multi-GPU training, diversity, smoothness, constant_noise, no_aug enables/disables the corresponding ablation setting as their name suggests. cfgs/task/single_train.yaml specifies the algorithm (DrQv2 or CURL) and the scene. cfgs/task/medium.yaml specifies the Data Steps and Linearly decayed noise schedule for DrQv2) cfgs/task/cma-es.yaml specifies cma-es specific hyperparameters.

To train GAIT-DrQ-v2 $ bash drqtrain.sh

To train GAIT-CURL $ bash curltrain.sh

To run CMA-ES $ bash cmatrain.sh

Use the corresponding watch script to watch the training output, e.g. $ bash drqwatch.sh

Evaluation

We provide 1 pre-trained checkpoint corresponding to DrQ-v2 default room0 located in logs/drqv2_habitat/66 16 multi 1m decay 3m train room_0_, containing only the actor network, and the training configurations.

To generate figures, sequence frames, and the corresponding interpolated videos in the paper with the pre-trained weights, first edit final_evaluation.py to choose which figure, then

$ python final_evaluation.py

Note that for training and evaluation, hydra will change the working directory to gait/logs/${algo}/${now:%Y.%m.%d}_${now:%H%M%S}.

Files

• rlcam_drqv2_mql.py and curl_train.py contains the main training code.
• drqv2_net.py contains the pytorch network architecture definitions for the Actor and the Critic networks.
• habitat_test.py contains our wrappers based on habitat-sim, which handles camera pose, computes reward, etc (Sections 3.1 and 3.2 of the paper.)
• drqv2/ and curl/ contains the original code from DrQ-v2 and CURL respectively. Our network architectures defined in drqv2_net.py are based on drqv2/drqv2.py.