bradyz / task-distillation

Code for Domain Adaptation Through Task Distillation (ECCV 20)

Domain Adaptation Through Task Distillation

Domain Adaptation Through Task Distillation
Brady Zhou*, Nimit Kalra*, Philipp Krähenbühl,
European Conference on Computer Vision (ECCV 2020)

@inproceedings{zhou2020domain,
    title={Domain Adaptation Through Task Distillation},
    author={Zhou, Brady and Kalra, Nimit and Kr{\"a}henb{\"u}hl, Philipp},
    booktitle={ECCV},
    year={2020}
}

Installation

• Grab a copy of our data (7 gb) and extract.
• Install the dependencies located in requirements.txt.
• OPTIONAL: To evaluate models (see them drive), download CARLA 0.9.6.

Training - Stage 1 (Proxy Model)

The first step involves training the proxy model, which is used to transfer a source model from VizDoom to CARLA.
The source model navigates a maze-like environment while avoiding poison bombs and we collect data that consists of rollouts generated by DFP.

The proxy model takes in a local semantic map representation (two channels, floor and poison) and outputs future waypoints.
This model only sees data from VizDoom.

python3 -m policy_transfer.train_stage1 --dataset_dir /home/bradyzhou/data/task-distillation-eccv20/vizdoom_data \\
    --source vizdoom \\
    --input_channels 2

This will generate a wandb dashboard with the training process.
Open it in your browser and you should see something like this -

A properly trained model will output predictions like this -

Training - Stage 2 (Target Model)

Next, we'll use this trained proxy model and distill it into a target model.
The target model takes in raw RGB images from CARLA as input and outputs future waypoints.
The future waypoints are generated by the proxy model, which now takes in the CARLA map representation as input (two channels, driveable area and obstacle).

Make sure the /path/to/proxy/ directory contains both the config.yaml config file, and model_*.t7 weights file.

python3 -m policy_transfer.train_image_v2.py --dataset_dir /home/bradyzhou/task-distillation-eccv20/carla_data \\
    --source carla \\
    --input_channel 3 \\
    --teacher_path /path/to/proxy/model_050.t7

This will generate a wandb dashboard dashboard with the training process.
Open it in your browser and you should see something like this -

A properly trained model will output predictions like this -

Evaluation

If you don't want to train from scratch, you can go to wandb project.

• Go to the stage-2 run
• Click on files
• Download model_020.t7 and config.yaml

Start the CARLA Server by going to the CARLA_0.9.6 directory and running

./CarlaUE4.sh -world-port=2000

Next, in another terminal window, add the CARLA .egg to your python path via

export PYTHONPATH=$PYTHONPATH:/path/to/CARLA_0.9.6/PythonAPI/carla/dist/carla-0.9.6-py3.5-linux-x86_64.egg

Then you can run the evaluation.
Make sure /path/to/weights/ contains both config.yaml and model_*.t7.

python3 -m policy_transfer.eval --agent_class ImageAgent \\
    --debug \\
    --port 2000 \\
    --pid 1 \\
    --agent_args /path/to/weights/model_020.t7