Skip to content

Latest commit

 

History

History
141 lines (122 loc) · 4.77 KB

README.md

File metadata and controls

141 lines (122 loc) · 4.77 KB

ViZDoom Experiments

This page contains the resources for the experiments of ViZDoom as discussed in the TLeague technical report.

Trained Model

A trained model for the experiments discussed in the technical report (CIG 2016 Track 1) has been given in the evaluation code, see the section below.

Evaluation

The evaluation code can be found here: Google Drive or Tencent Weiyun, which is a modification over the original evaluation code https://github.com/mihahauke/VDAIC2017

Our modification allows a synchronous mode for the host, "F1" and "MyPlayer", as discussed in the technical report and is summarized below.

The root config file _vizdoom.cfg overrides all the private configurations of an agent, so we've commented out the ASYNC_PLAYER setting in _vizdoom.cfg:

doom_scenario_path = cig2017.wad
# window_visible = False
# mode = ASYNC_PLAYER
game_args += -join localhost

For "F1", we add extra arguments to f1/my_glorious_agent.py

    -c F1_COLOR, --color F1_COLOR
                            0 - green, 1 - gray, 2 - brown, 3 - red, 4 - light
                            gray, 5 - light brown, 6 - light red, 7 - light blue
                            (default: 0)
    -w, --watch           window visible (default: False)
    -mode MODE, --mode MODE
                            1 for PLAYER, 2 for ASYNC_PLAYER (default: 1)

and default "F1" to green color and synchronous player.

For the host, we add an argument

    -mode MODE, --mode MODE
                            1 for PLAYER, 2 for ASYNC_PLAYER, 3 for SPECTATOR, 4
                            for ASYNC_SPECTATOR (default: 1)

to allow it be synchronous (mode 3, the SPECTATOR).

Installation

The official code requires each submitted agent be packed as docker (see the descriptions at https://github.com/mihahauke/VDAIC2017). In our modified code, we do a common pip install for "MyPlayer", and docker build for any other third-party agent (e.g., "F1"), as explained below.

Install MyPlayer

Install the three packages Arena, TLeague, TPolicies, respectively:

cd MyPlayer/Arena
pip3 install -e .
cd MyPlayer/TLeague
pip3 install -e .
cd MyPlayer/TPolicies
pip3 install -e .

Note, the MyPlayer evaluation code here is self-contained and relies on the old TLeague, Arena, TPolicies code that is somehow different from the dev-open branch. You can avoid the possible conflicts by using, for example, virtualenv.

An NN model has been shipped with the code and placed in the path MyPlayer/model/.

Install F1

Build the image:

sudo chmod u+x build.sh
./build.sh f1

The corresponding NN model has been contained.

Install host

Build the image:

sudo chmod u+x build.sh
./build.sh host

Run

To run the evaluation, start the host, MyPlayer and F1 in separate terminals. tmux is recommended. Also, ensure

sudo chmod u+x run.sh

For 1 MyPlayer, 7 builtin bots, run the following commands in separate terminals:

./run.sh host -mode 3 -b 7 
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation

where -mode 3 means synchronous spectator, -b 7 means adding 7 builtin bots.

For 1 MyPlayer, 1 F1 and 6 builtin bots, run the following commands in separate terminals:

./run.sh host -mode 3 -b 6 -p 2
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation
./run.sh f1

where -p 2 means there are 2 AI agents to join in.

For 4 MyPlayer, 4 F1, run the following commands in separate terminals:

./run.sh host -mode 3 -p 8
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation
bash MyPlayer/TLeague/tleague/sandbox/example_evaluation_vd.sh evaluation
./run.sh f1
./run.sh f1
./run.sh f1
./run.sh f1

Training Code

As described in the technical reports, there are two stages for training: one for the navigation, and the other for the frag. We provide the corresponding .yml.jinja2 files here: for navigation and for frag, respectively.

Run the training over a k8s cluster:

# start
python render_template.py vdtr-navi-open.yml.jinja2 | kubectl apply -f -
# stop
python render_template.py vdtr-navi-open.yml.jinja2 | kubectl delete -f -
# start
python render_template.py vdtr-frag-open.yml.jinja2 | kubectl apply -f -
# stop
python render_template.py vdtr-frag-open.yml.jinja2 | kubectl delete -f -

TODO: guidance to setting up PVC?

Downloads

TODO: link to the video clips for the evaluation