This is an implementation of a neural-network based Go AI, using TensorFlow. While inspired by DeepMind's AlphaGo algorithm, this project is not a DeepMind project nor is it affiliated with the official AlphaGo project.
Repeat, this is not the official AlphaGo program by DeepMind. This is an independent effort by Go enthusiasts to replicate the results of the AlphaGo Zero paper ("Mastering the Game of Go without Human Knowledge," Nature), with some resources generously made available by Google.
Minigo is based off of Brian Lee's "MuGo" -- a pure Python implementation of the first AlphaGo paper "Mastering the Game of Go with Deep Neural Networks and Tree Search" published in Nature. This implementation adds features and architecture changes present in the more recent AlphaGo Zero paper, "Mastering the Game of Go without Human Knowledge". More recently, this architecture was extended for Chess and Shogi in "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm". These papers will often be abridged in Minigo documentation as AG (for AlphaGo), AGZ (for AlphaGo Zero), and AZ (for AlphaZero) respectively.
-
Provide a clear set of learning examples using Tensorflow, Kubernetes, and Google Cloud Platform for establishing Reinforcement Learning pipelines on various hardware accelerators.
-
Reproduce the methods of the original DeepMind AlphaGo papers as faithfully as possible, through an open-source implementation and open-source pipeline tools.
-
Provide our data, results, and discoveries in the open to benefit the Go, machine learning, and Kubernetes communities.
An explicit non-goal of the project is to produce a competitive Go program that establishes itself as the top Go AI. Instead, we strive for a readable, understandable implementation that can benefit the community, even if that means our implementation is not as fast or efficient as possible.
While this product might produce such a strong model, we hope to focus on the process. Remember, getting there is half the fun. :)
We hope this project is an accessible way for interested developers to have access to a strong Go model with an easy-to-understand platform of python code available for extension, adaptation, etc.
If you'd like to read about our experiences training models, see RESULTS.md.
To see our guidelines for contributing, see CONTRIBUTING.md.
This project assumes you have the following:
The Hitchhiker's guide to python has a good intro to python development and virtualenv usage. The instructions after this point haven't been tested in environments that are not using virtualenv.
pip3 install virtualenv
pip3 install virtualenvwrapper
wget https://github.com/bazelbuild/bazel/releases/download/0.19.2/bazel-0.19.2-installer-linux-x86_64.sh
chmod 755 bazel-0.19.2-installer-linux-x86_64.sh
sudo ./bazel-0.19.2-installer-linux-x86_64.sh
rm bazel-0.19.2-installer-linux-x86_64.sh
First set up and enter your virtualenv and then the shared requirements:
pip3 install -r requirements.txt
Then, you'll need to choose to install the GPU or CPU tensorflow requirements:
- GPU:
pip3 install "tensorflow-gpu>=1.13,<1.14"
.- Note: You must install CUDA 9.0. for Tensorflow 1.5+.
- CPU:
pip3 install "tensorflow>=1.13,<1.14"
.
You may want to use a cloud project for resources. If so set:
PROJECT=foo-project
Then, running
source cluster/common.sh
will set up other environment variables defaults.
./test.sh
To run individual modules
BOARD_SIZE=9 python3 tests/run_tests.py test_go
BOARD_SIZE=19 python3 tests/run_tests.py test_mcts
To automatically test PRs, Minigo uses Prow, which is a test framework created by the Kubernetes team for testing changes in a hermetic environment. We use prow for running unit tests, linting our code, and launching our test Minigo Kubernetes clusters.
You can see the status of our automated tests by looking at the Prow and Testgrid UIs:
- Testgrid (Test Results Dashboard): https://k8s-testgrid.appspot.com/sig-big-data
- Prow (Test-runner dashboard): https://prow.k8s.io/?repo=tensorflow%2Fminigo
All commands are compatible with either Google Cloud Storage as a remote file system, or your local file system. The examples here use GCS, but local file paths will work just as well.
To use GCS, set the BUCKET_NAME
variable and authenticate via gcloud login
.
Otherwise, all commands fetching files from GCS will hang.
For instance, this would set a bucket, authenticate, and then look for the most recent model.
# When you first start we recommend using our minigo-pub bucket.
# Later you can setup your own bucket and store data there.
export BUCKET_NAME=minigo-pub/v9-19x19
gcloud auth application-default login
gsutil ls gs://$BUCKET_NAME/models | tail -4
Which might look like:
gs://$BUCKET_NAME/models/000737-fury.data-00000-of-00001
gs://$BUCKET_NAME/models/000737-fury.index
gs://$BUCKET_NAME/models/000737-fury.meta
gs://$BUCKET_NAME/models/000737-fury.pb
These four files comprise the model. Commands that take a model as an
argument usually need the path to the model basename, e.g.
gs://$BUCKET_NAME/models/000737-fury
You'll need to copy them to your local disk. This fragment copies the files
associated with $MODEL_NAME
to the directory specified by MINIGO_MODELS
:
MODEL_NAME=000737-fury
MINIGO_MODELS=$HOME/minigo-models
mkdir -p $MINIGO_MODELS/models
gsutil ls gs://$BUCKET_NAME/models/$MODEL_NAME.* | \
gsutil cp -I $MINIGO_MODELS/models
To watch Minigo play a game, you need to specify a model. Here's an example to play using the latest model in your bucket
python3 selfplay.py \
--verbose=2 \
--num_readouts=400 \
--load_file=$MINIGO_MODELS/models/$MODEL_NAME
where READOUTS
is how many searches to make per move. Timing information and
statistics will be printed at each move. Setting verbosity to 3 or
higher will print a board at each move.
Minigo uses the GTP Protocol, and you can use any gtp-compliant program with it.
# Latest model should look like: /path/to/models/000123-something
LATEST_MODEL=$(ls -d $MINIGO_MODELS/* | tail -1 | cut -f 1 -d '.')
python3 gtp.py --load_file=$LATEST_MODEL --num_readouts=$READOUTS --verbose=3
After some loading messages, it will display GTP engine ready
, at which point
it can receive commands. GTP cheatsheet:
genmove [color] # Asks the engine to generate a move for a side
play [color] [coordinate] # Tells the engine that a move should be played for `color` at `coordinate`
showboard # Asks the engine to print the board.
One way to play via GTP is to use gogui-display (which implements a UI that speaks GTP.) You can download the gogui set of tools at http://gogui.sourceforge.net/. See also documentation on interesting ways to use GTP.
gogui-twogtp -black 'python3 gtp.py --load_file=$LATEST_MODEL' -white 'gogui-display' -size 19 -komi 7.5 -verbose -auto
Another way to play via GTP is to watch it play against GnuGo, while spectating the games:
BLACK="gnugo --mode gtp"
WHITE="python3 gtp.py --load_file=$LATEST_MODEL"
TWOGTP="gogui-twogtp -black \"$BLACK\" -white \"$WHITE\" -games 10 \
-size 19 -alternate -sgffile gnugo"
gogui -size 19 -program "$TWOGTP" -computer-both -auto
The following sequence of commands will allow you to do one iteration of reinforcement learning on 9x9. These are the basic commands used to produce the models and games referenced above.
The commands are
- bootstrap: initializes a random model
- selfplay: plays games with the latest model, producing data used for training
- train: trains a new model with the selfplay results from the most recent N generations.
Training works via tf.Estimator; a working directory manages checkpoints and training logs, and the latest checkpoint is periodically exported to GCS, where it gets picked up by selfplay workers.
Configuration for things like "where do debug SGFs get written", "where does training data get written", "where do the latest models get published" are managed by the helper scripts in the rl_loop directory. Those helper scripts execute the same commands as demonstrated below. Configuration for things like "what size network is being used?" or "how many readouts during selfplay" can be passed in as flags. The mask_flags.py utility helps ensure all parts of the pipeline are using the same network configuration.
All local paths in the examples can be replaced with gs://
GCS paths, and the
Kubernetes-orchestrated version of the reinforcement learning loop uses GCS.
This command initializes your working directory for the trainer and a random
model. This random model is also exported to --model-save-path
so that
selfplay can immediately start playing with this random model.
If these directories don't exist, bootstrap will create them for you.
export MODEL_NAME=000000-bootstrap
python3 bootstrap.py \
--work_dir=estimator_working_dir \
--export_path=outputs/models/$MODEL_NAME
This command starts self-playing, outputting its raw game data as tf.Examples as well as in SGF form in the directories.
python3 selfplay.py \
--load_file=outputs/models/$MODEL_NAME \
--num_readouts 10 \
--verbose 3 \
--selfplay_dir=outputs/data/selfplay \
--holdout_dir=outputs/data/holdout \
--sgf_dir=outputs/sgf
This command takes a directory of tf.Example files from selfplay and trains a
new model, starting from the latest model weights in the estimator_working_dir
parameter.
Run the training job:
python3 train.py \
outputs/data/selfplay/* \
--work_dir=estimator_working_dir \
--export_path=outputs/models/000001-first_generation
At the end of training, the latest checkpoint will be exported to. Additionally, you can follow along with the training progress with TensorBoard. If you point TensorBoard at the estimator working directory, it will find the training log files and display them.
tensorboard --logdir=estimator_working_dir
It can be useful to set aside some games to use as a 'validation set' for
tracking the model overfitting. One way to do this is with the validate
command.
By default, Minigo will hold out 5% of selfplay games for validation. This can
be changed by adjusting the holdout_pct
flag on the selfplay
command.
With this setup, rl_loop/train_and_validate.py
will validate on the same
window of games that were used to train, writing TensorBoard logs to the
estimator working directory.
This might be useful if you have some known set of 'good data' to test your network against, e.g., a set of pro games. Assuming you've got a set of .sgfs with the proper komi & boardsizes, you'll want to preprocess them into the .tfrecord files, by running something similar to
import preprocessing
filenames = [generate a list of filenames here]
for f in filenames:
try:
preprocessing.make_dataset_from_sgf(f, f.replace(".sgf", ".tfrecord.zz"))
except:
print(f)
Once you've collected all the files in a directory, producing validation is as easy as
python3 validate.py \
validation_files/ \
--work_dir=estimator_working_dir \
--validation_name=pro_dataset
The validate.py will glob all the .tfrecord.zz files under the directories given as positional arguments and compute the validation error for the positions from those files.
See more at cluster/README.md