Chinese Standard Mahjong

RL Framework: tianshou

• learning
  • imitation_trainer.py: trainer, see test/test_imitator.py or imitate.py
  • imitation.py: policy of imitation learning wrapped for tianshou, see test/test_imitator.py or imitate.py
  • model.py: models of resnet
  • ppo.py: not used
  • wrapper.py: policy of neural network wrapped for env.runner.Runner, see imitate.py or scripts/botzone.py

None-Learning Wrapper

• env

  • bot.py: not used
  • imitator.py: wrapped environment for imitation learning, see imitate.py
  • multirunner.py: wrapped parallel environment for Mahjong, see test/test_multirunner.py or scripts/perf.py
  • runner.py: wrapped environment for Mahjong, see scripts/example.py

• utils

  • distribution.py: for utils.policy.random_policy, see utils/policy.py
  • match_data.py: dump actions of expert data for imitation learning, see scripts/preprocess.py
  • merge.py: offer decorator to package the code files, see scripts/botzone.py
  • paired_data.py: dump observation-action pairs from match data for imitation learning, see scripts/generate_pair.py
  • policy.py: implentation of some wrapped policy
  • sync_buffer.py: a buffer for multi-thread control, see env/multirunner.py
  • tile_traits.py: wrapper of player information to offer one-hot observation
  • vec_data.py: wrapper of player information to offer one-hot observation

• scripts

  • botzone.py: botzone interface
    • python -m scripts.botzone -p data/best.pth -bn --merge
  • example.py: examples of some API
    • python -m scripts.example
  • generate_pair.py: generate observation-action pairs from match data
    • python -m scripts.generate_pair match.pkl -o pair.pkl
  • perf.py: evaluate performance of a certain model
    • python -m scripts.perf -p A.pth -cp B.pth -bn
  • preprocess.py: generate match data from match log (old version)
    • python -m scripts.preprocess data.txt -o match.pkl

• test

  • test_augmentation.py: check whether the augmentation is correctly implemented
    • python -m test.test_augmentation pair.pkl
  • test_imitator.py: examples of imitation learning with match data
    • python -m test.test_imitator -f match.pkl -d train -e experiment1
  • test_multirunner.py: check whether the multi-thread runner works correctly
    • python -m test.test_multirunner
  • test_paired_data.py: check whether the encoder and decoder in utils/paired_data.py work correctly
    • python -m test.test_paired_data match.pkl

• deprecated

  • some deprecated scripts (most of which are for performance reasons)

Usage

Train

python -m scripts.preprocess data.txt -o match.pkl
python -m scripts.generate_pair match.pkl -o pair.pkl
python imitate.py -f pair.pkl -d train -e experiment1

Visualization

python imitate.py -f pair.pkl -d train -e experiment1 --eval

Evaluate

python -m scripts.perf -p data/best.pth -cp data/cmp.pth -bn

Upload

python -m scripts.botzone -p data/best.pth -bn --merge

• Then upload the merge.zip as well as your best.pth onto the Botzone.