This repository provides an example of how to integrate a solution based on Behavioral Cloning into the submission kit for the 2021 MineRL BASALT competition.
MineRL BASALT is a competition on solving human-judged tasks. The tasks in this competition do not have a pre-defined reward function: the goal is to produce trajectories that are judged by real humans to be effective at solving a given task.
See the homepage of the competition for further details.
This repository contains:
- Documentation on how to submit your agent to the leaderboard
- The procedure for Round 1 and Round 2
- Starter code that uses the imitation implementation of Behavioral Cloning to train a simple agent.
Other Resources:
- AICrowd competition page - Main registration page & leaderboard.
- MineRL Documentation - Documentation for the
minerl
package! - Imitation Documentation - Documentation for the
imitation
package, which trains imitation-based models in the style of Stable Baselines 3 - Sacred Documentation - Documentation for the
sacred
package, which is used to structure and define configurations for experiments
This section of the repo is structured as a small utility package, containing wrappers, tools, and compatibility wrappers
that allow us to more easily train on top of BASALT environments. It should be automatically installed by following the
setup instructions and originating a conda environment from environment.yml
, but can also be installed manually by navigating
into the directory and calling pip install .
This section of the repo is where the actual logic of a Behavioral Cloning training procedure is laid out, specifically in
basalt_baselines/bc.py
; other baselines are current works in progress.
The training procedure is structure as a Sacred experiment. The most salient things to know about this are:
- Configuration values specified in the method decorated with
@bc_baseline.config
(in this case,default_config
), are automatically made available to any functions decorated with@bc_baseline.capture
,@bc_baseline.main
, orbc_baseline.automain
. - If you want to run the testing or training code directly, you can call
basalt_baselines/bc.py with mode='train'
orwith mode='test'
. You can also experiment with different configuration parameters on the command line by specifying a new value of anything defined in the config method (mentioned in (1)). For example, you could callbasalt_baselines.bc.py with batch_size=16
This BC baseline is meant to be simple and minimal, and, as such, it tries to make the simplest design choices that allow it to handle the structure of a Minecraft environment. These include:
- Extracting only the pixel POV observation, and using a CNN on that observation as the input to the BC model
- Turning the continuous camera action into discretely chunked left/right and up/down movements, since otherwise the scale of the log likelihood for the continuous space swamps the discrete spaces
- Constructing separate action distributions for each of the actions that make up a joint Minecraft action (which can be a mix of Discrete and Box), and combining those action distributions into one MultiModalActionDistribution that is used to predict the actions being predicted for BC (by sampling independently from each action space's distribution)
- This action distribution architecture works by learning a single latent vector, and then feeding that representation to a head mapping it into the parameters required by each action's distribution
In brief: you define your Python environment using Anaconda environment files, and AICrowd system will build a Docker image and run your code using the docker scripts inside the utility
directory.
You submit pretrained models, the evaluation code and the training code. Training code should produce the same models you upload as part of your submission.
Your evaluation code (test_submission_code.py
) only needs to control the agent and accomplish the environment's task. The evaluation server will handle recording of videos.
You specify tasks you want to submit agent for with aicrowd.json
file, tags
field (see below).
-
Clone the github repository or press the "Use this Template" button on GitHub!
git clone https://github.com/minerllabs/basalt_competition_submission_template.git
-
Install the Java JDK! Make sure you have the JDK 8 installed first!
-> Go to http://minerl.io/docs/tutorials/getting_started.html
-
Specify your specific submission dependencies (PyTorch, Tensorflow, etc)
- Anaconda Environment. To run this baseline code on your local machine, you will need to
create an environment with the correct dependencies on your local machine. We recommend
anaconda
for this purpose, and have included anenvironment.yml
file specifying necessary dependencies to run our BC baseline. Make sure at least version4.5.11
ofanaconda
is installed (By following instructions here).
Also, if you are not on a machine with NVIDIA drivers that can support
cudatoolkit=10.2
, remove that dependency from theenvironment.yml
file before trying to installThen: * Create your new conda environment
```sh conda-env create -f environment.yml conda activate basalt ``` This will install the `minerl` environment (containing all of the competition environments), as well as dependencies used in the training of the baselines themselves.
-
Your code specific dependencies Add your own dependencies to the
environment.yml
file. Remember to add any additional channels. PyTorch requires the channelpytorch
, for example. You can also install them locally usingconda install <your-package>
-
Pip Packages If you need pip packages (not on conda), you can add them to the
environment.yml
file (see the currently populated version): -
Apt Packages If your training procedure or agent depends on specific Debian (Ubuntu, etc.) packages, add them to
apt.txt
.
- Anaconda Environment. To run this baseline code on your local machine, you will need to
create an environment with the correct dependencies on your local machine. We recommend
These files are used to construct both the local and AICrowd docker containers in which your agent will train.
If above are too restrictive for defining your environment, see this Discourse topic for more information.
Please follow the example structure shared in the starter kit for the code structure. The different files and directories have following meaning:
.
├── aicrowd.json # Submission meta information like your username
├── apt.txt # Packages to be installed inside docker image
├── data # The downloaded data, the path to directory is also available as `MINERL_DATA_ROOT` env variable
├── test_submission_code.py # IMPORTANT: Your testing/inference phase code. NOTE: This is NOT the the entry point for testing phase!
├── train # Your trained model MUST be saved inside this directory
├── train_submission_code.py # IMPORTANT: Your training code. Running this should produce the same agent as you upload as part of the agent.
├── test_framework.py # The entry point for the testing phase, which sets up the environment. Your code DOES NOT go here.
└── utility # The utility scripts which provide a smoother experience to you.
├── debug_build.sh
├── docker_run.sh
├── environ.sh
├── evaluation_locally.sh
├── parser.py
├── train_locally.sh
└── verify_or_download_data.sh
Finally, you must specify an AIcrowd submission JSON in aicrowd.json
to be scored!
The aicrowd.json
of each submission should contain the following content:
{
"challenge_id": "neurips-2021-minerl-basalt-competition",
"authors": ["your-aicrowd-username"],
"description": "sample description about your awesome agent",
"tags": ["FindCave", "MakeWaterfall", "CreateVillageAnimalPen", "BuildVillageHouse"],
"license": "MIT",
"gpu": true
}
This JSON is used to map your submission to the said challenge, so please remember to use the correct challenge_id
as specified above.
You can specify which tasks you want to be re-evaluated with the tags
field by specifying the list of tags. If a task is not included in this list, your current submission for that task will not be updated.
Please specify if your code will use a GPU or not for the evaluation of your model. If you specify true
for the GPU, a NVIDIA Tesla K80 GPU will be provided and used for the evaluation.
You don't need to upload the MineRL dataset in submission and it will be provided in online submissions at MINERL_DATA_ROOT
path, should you need it. For local training and evaluations, you can download it once in your system via python ./utility/verify_or_download_data.py
or place manually into the ./data/
folder.
To make a submission, you will have to create a private repository on https://gitlab.aicrowd.com/.
You will have to add your SSH Keys to your GitLab account by following the instructions here. If you do not have SSH Keys, you will first need to generate one.
Then you can create a submission by making a tag push to your repository on https://gitlab.aicrowd.com/.
Any tag push (where the tag name begins with "submission-") to your private repository is considered as a submission
Then you can add the correct git remote, and finally submit by doing :
cd competition_submission_starter_template
# Add AIcrowd git remote endpoint
git remote add aicrowd [email protected]:<YOUR_AICROWD_USER_NAME>/basalt_competition_submission_template.git
git push aicrowd master
# Create a tag for your submission and push
git tag -am "submission-v0.1" submission-v0.1
git push aicrowd master
git push aicrowd submission-v0.1
# Note : If the contents of your repository (latest commit hash) does not change,
# then pushing a new tag will **not** trigger a new evaluation.
You now should be able to see the details of your submission at: https://gitlab.aicrowd.com/<YOUR_AICROWD_USER_NAME>/basalt_competition_submission_template/issues/
Best of Luck 🎉 🎉
You can perform local training and evaluation using utility scripts shared in this directory. To mimic the online training phase you can run ./utility/train_locally.sh
from the repository root, you can specify --verbose
for complete logs.
For local evaluation of your code, you can use ./utility/evaluation_locally.sh
, add --verbose
if you want to view complete logs. Note that you do not need to record videos in your code! AICrowd server will handle this. Your code only needs to play the games.
For running/testing your submission in a docker environment (identical to the online submission), you can use ./utility/docker_train_locally.sh
and ./utility/docker_evaluation_locally.sh
. You can also run docker image with bash entrypoint for debugging on the go with the help of ./utility/docker_run.sh
. These scripts respect following parameters:
--no-build
: To skip docker image build and use the last build image--nvidia
: To usenvidia-docker
instead ofdocker
which include your nvidia related drivers inside docker image
The quick-start kit was authored by Anssi Kanervisto and Shivam Khandelwal with help from William H. Guss
The BASALT competition is organized by the following team:
- Rohin Shah (UC Berkeley)
- Cody Wild (UC Berkeley)
- Steven H. Wang (UC Berkeley)
- Neel Alex (UC Berkeley)
- Brandon Houghton (OpenAI and Carnegie Mellon University)
- William H. Guss (OpenAI and Carnegie Mellon University)
- Sharada Mohanty (AIcrowd)
- Anssi Kanervisto (University of Eastern Finland)
- Stephanie Milani (Carnegie Mellon University)
- Nicholay Topin (Carnegie Mellon University)
- Pieter Abbeel (UC Berkeley)
- Stuart Russell (UC Berkeley)
- Anca Dragan (UC Berkeley)