search

so-much-meta / lczero_tools

Utilities for experimenting with leela-chess-zero
GNU General Public License v3.0
42 stars 22 forks source link

readme

lczero-tools

Python utilities for experimenting with Leela Chess Zero a neural network based chess engine: https://github.com/glinscott/leela-chess/

IMPORTANT - DEFUNCT / NO LONGER BEING MAINTAINED

Please go to https://github.com/LeelaChessZero/lc0/tree/master/scripts/pybind for the official python bindings!

I have not actively worked on this in years. Many things have changed to Leela Chess Zero's network architecture in the meantime, including a different weights file format. I have intended to come back to this to update it, and still would like to, but for now this project is where it is. This will not work with the latest versions of Leela's network without significant changes. However, maybe somebody sees something useful here or wants to update this project.

Note: This is primarily for looking at the Leela Chess neural network itself, outside of search/MCTS (although search may be added eventually).

This makes heavy use of python-chess located at https://github.com/niklasf/python-chess

The current implementation is primarily geared towards pytorch, but tensorflow is possible using the training/tf portion of leela-chess.

Example usage (also see /tests/*.py and Examples.ipynb):

>>> from lcztools import load_network, LeelaBoard
>>> # Note: use pytorch_cuda for cuda support
>>> net = load_network('pytorch_cuda', 'weights.txt.gz')
>>> board = LeelaBoard()
>>> # Many of Python-chess's methods are passed through, along with board representation
>>> board.push_uci('e2e4')
>>> print(board)
r n b q k b n r
p p p p p p p p
. . . . . . . .
. . . . . . . .
. . . . P . . .
. . . . . . . .
P P P P . P P P
R N B Q K B N R
Turn: Black
>>> policy, value = net.evaluate(board)
>>> print(policy)
OrderedDict([('c7c5', 0.5102739), ('e7e5', 0.16549255), ('e7e6', 0.11846365), ('c7c6', 0.034872748),
('d7d6', 0.025344277), ('a7a6', 0.02313047), ('g8f6', 0.021814445), ('g7g6', 0.01614216), ('b8c6', 0.013772337),
('h7h6', 0.013361011), ('b7b6', 0.01300134), ('d7d5', 0.010980369), ('a7a5', 0.008497312), ('b8a6', 0.0048270077),
('g8h6', 0.004309486), ('f7f6', 0.0040882644), ('h7h5', 0.003910391), ('b7b5', 0.0027878743), ('f7f5', 0.0025032777),
('g7g5', 0.0024271626)])
>>> print(value)
0.4715215042233467

Create network server

It is possible to load the network (or multiple different networks) once in a network server, and access this by multiple clients. This does not add much overhead, and creates a significant speedup by batching GPU operations if multiple clients are simultaneously connected.

IPC communication is via zeromq.

python -m lcztools.backend.net_server.server weights_file1.txt.gz weights_file2.txt.gz

After the server starts, clients can access it like so using the load_network interface:

>>> from lcztools import load_network, LeelaBoard
>>> net0 = load_network(backend='net_client', network_id=0)
>>> net1 = load_network(backend='net_client', network_id=1)
>>> board = LeelaBoard()
>>> policy0, value0 = net0.evaluate(board)
>>> policy1, value1 = net1.evaluate(board)

Max batch size can be configured by entering it after the weights file. Default is 32. Batch sizes are generally powers of 2 (starting at 1), and it can help to set this to the batch size that will actually be used if less than 32 clients are connected. When a new client connects, it sends a "hi" message to the server, which causes the server to reset the batch size to the max_batch_size (this message can also be sent throughout a program's executin via net.hi()). The network server will block up to 1 second if the batch is not filled, at which time the batch size is reset to the greatest power of 2 less than or equal to the number of items currently in the batch.

python -m lcztools.backend.net_server.server weights_file1.txt.gz 8 weights_file2.txt.gz 8

INSTALL

# With both torch and util dependencies for NN evaluation
pip install git+https://github.com/so-much-meta/lczero_tools.git#egg=lczero-tools[torch,util]
# Or just util extras (parse training games, run lczero engine, etc)
pip install git+https://github.com/so-much-meta/lczero_tools.git#egg=lczero-tools[util]

# Or from source tree...
git clone https://github.com/so-much-meta/lczero_tools
cd lczero_tools
# Note: Creating and using a virtualenv or Conda environment before install is suggested, as always
pip install .[torch,util]
# Or for developer/editable install, to make in place changes:
# pip install -e .[torch,util]

TODO

  1. [x] DONE: Implement testing to verify position evaluations match lczero engine.
    • [ ] Using /tests/test_net_eq_engine.py, results look good. But specific PGNs might be helpful too.
  2. [x] DONE: Add config mechanism and Jupyter notebook examples
  3. [x] DONE: Add training data parser module. Use cases are:
    • [x] DONE: Training data to PGN
    • [ ] Verification of training data correctness.
    • [ ] Loss calculation - allow comparison between networks on same data
  4. [x] DONE: lczero web scraping (NOT FOR HEAVY USE)
    • [x] DONE: Convert individidual match and training games to PGN (URL => PGN)
    • [x] DONE: Download weights files
  5. [ ] OpenCL support! This should be possible with https://github.com/plaidml/plaidml
  6. [ ] Investigate optimizations (CUDA, multiprocessing, etc). Goal is to eventually have a fast enough python-based implementation to do MCTS and get decent nodes/second comparable to Leela's engine -- in cases where neural network eval speed is the bottleneck.
    • [ ] However, no optimizations should get (too much) in the way of clarity or ease of changing code to do experiments.
  7. [ ] Possible MCTS implementation