def __init__(self, board_size=BOARD_SIZE, color=BLACK, board=None):
'''
Args:
board_size: size of board
color: color of current player
board: current board
'''
assert color in [BLACK, WHITE], 'Invalid player color'
if board:
self.board = board
else:
self.board = pachi_py.CreateBoard(board_size)
self.last_action = -1
self.last_action_2 = -1
self.color = color
self.board_size = board_size
self.game_over = False
self.winner = None
self.action_space = spaces.Discrete(board_size**2 + 1)
self._new_state_checks()
if color == BLACK:
self.current_player = 1
else:
self.current_player = -1
def reset(self):
""" Reset the board """
self.board = pachi_py.CreateBoard(self.board_size)
opponent_resigned = False
self.done = self.board.is_terminal or opponent_resigned
return _format_state(self.history, self.player_color, self.board_size)
def __init__(self,
board_size,
board=None,
player=-1,
done=False,
last_passed=False,
history=None,
move_num=0):
self.board_size = board_size
self.pass_action = board_size**2
self.resign_action = board_size**2 + 1
assert player in [-1, 1]
self.curr_player = player
self.done = done
self.last_passed = last_passed
self.move_num = move_num
if history is None:
self.history = [None] * 7
else:
self.history = history
if board is None:
self.board = pachi_py.CreateBoard(board_size)
else:
self.board = board
def make_random_board(size):
b = pachi_py.CreateBoard(size)
c = pachi_py.BLACK
for _ in range(0, 50):
b = b.play(np.random.choice(b.get_legal_coords(c)), c)
c = pachi_py.stone_other(c)
return b
def test_illegal_move():
b = pachi_py.CreateBoard(9).play(14, pachi_py.WHITE)
try:
b.play(14, pachi_py.BLACK)
except pachi_py.IllegalMove:
return
assert False, 'IllegalMove exception should have been raised'
def __init__(self, player_color, observation_type, illegal_move_mode,
board_size, komi):
"""
Args:
player_color: Stone color for the agent. Either 'black' or 'white'
opponent: An opponent policy
observation_type: State encoding
illegal_move_mode: What to do when the agent makes an illegal move. Choices: 'raise' or 'lose'
"""
assert isinstance(
board_size,
int) and board_size >= 1, 'Invalid board size: {}'.format(
board_size)
self.board_size = board_size
self.komi = komi
self._seed()
colormap = {
'black': pachi_py.BLACK,
'white': pachi_py.WHITE,
}
try:
self.player_color = colormap[player_color]
except KeyError:
raise error.Error(
"player_color must be 'black' or 'white', not {}".format(
player_color))
# self.opponent_policy = None
# self.opponent = opponent
assert observation_type in ['image3c']
self.observation_type = observation_type
assert illegal_move_mode in ['lose', 'raise']
self.illegal_move_mode = illegal_move_mode
if self.observation_type != 'image3c':
raise error.Error('Unsupported observation type: {}'.format(
self.observation_type))
shape = pachi_py.CreateBoard(self.board_size).encode().shape
self.observation_space = spaces.Box(np.zeros(shape), np.ones(shape))
# One action for each board position, pass, and resign
self.action_space = spaces.Discrete(self.board_size**2 + 2)
# Filled in by _reset()
self.state = None
self.done = True
# Modifications
self.last_player_passed = False
self.BLACK = np.array([1, 0, 0])
self.WHITE = np.array([0, 1, 0])
self.EMPTY = np.array([0, 0, 1])
def reset(self):
self.state = GoState(pachi_py.CreateBoard(self.board_size), pachi_py.BLACK)
# (re-initialize) the opponent
# necessary because a pachi engine is attached to a game via internal data in a board
# so with a fresh game, we need a fresh engine
# self._reset_opponent(self.state.board)
# Let the opponent play if it's not the agent's turn
opponent_resigned = False
# We should be back to the agent color
self.last_player_passed = False
self.done = self.state.board.is_terminal
return self.state.board.encode()
def _seed(self, seed=None):
self.np_random, seed1 = seeding.np_random(seed)
# Derive a random seed.
seed2 = seeding.hash_seed(seed1 + 1) % 2**32
pachi_py.pachi_srand(seed2)
shape = pachi_py.CreateBoard(self.board_size).encode().shape
self.observation_space = spaces.Box(np.zeros(shape),
np.ones(shape),
np_random=self.np_random)
# One action for each board position, pass, and resign
self.action_space = spaces.Discrete(self.board_size**2 + 2,
np_random=self.np_random)
return [seed1, seed2]
def _reset(self):
self.state = GoState(pachi_py.CreateBoard(self.board_size), pachi_py.BLACK)
# (re-initialize) the opponent
# necessary because a pachi engine is attached to a game via internal data in a board
# so with a fresh game, we need a fresh engine
self._reset_opponent(self.state.board)
# Let the opponent play if it's not the agent's turn
if self.state.color != self.player_color:
self.state = self._exec_opponent_play(self.state, None, None)
assert self.state.color == self.player_color
self.done = self.state.board.is_terminal
return self.state.board.encode()
def reset(self):
""" Reset the board """
colormap = {
'black': pachi_py.BLACK,
'white': pachi_py.WHITE,
}
self.player_color = colormap['black']
self.history = [np.zeros((HISTORY + 1, self.board_size, self.board_size)),
np.zeros((HISTORY + 1, self.board_size, self.board_size))]
self.board = pachi_py.CreateBoard(self.board_size)
opponent_resigned = False
self.state = _format_state(self.history,
self.player_color, self.board_size)
self.done = self.board.is_terminal or opponent_resigned
return self.state
def __init__(self, player_color, opponent, observation_type,
illegal_move_mode, board_size):
'''
Args:
player_color: Stone color for the agent. Either 'black' or 'white'
opponent: An opponent policy
observation_type: State encoding
illegal_move_mode: What to do when the agent makes an illegal move. Choices: 'raise' or 'lose'
'''
assert isinstance(
board_size,
int) and board_size >= 1, 'Invalid board size: {}'.format(
board_size)
self.board_size = board_size
colormap = {
'black': pachi_py.BLACK,
'white': pachi_py.WHITE,
}
try:
self.player_color = colormap[player_color]
except KeyError:
raise error.Error(
"player_color must be 'black' or 'white', not {}".format(
player_color))
self.opponent_policy = None
self.opponent = opponent
assert observation_type in ['image3c']
self.observation_type = observation_type
assert illegal_move_mode in ['lose', 'raise']
self.illegal_move_mode = illegal_move_mode
# One action for each board position, pass, and resign
self.action_space = spaces.Discrete(self.board_size**2 + 2)
if self.observation_type == 'image3c':
shape = pachi_py.CreateBoard(self.board_size).encode().shape
self.observation_space = spaces.Box(np.zeros(shape),
np.ones(shape))
else:
raise error.Error('Unsupported observation type: {}'.format(
self.observation_type))
self.reset()
def _play(black_policy_fn, white_policy_fn, board_size=19):
'''
Samples a trajectory for two player policies.
Args:
black_policy_fn, white_policy_fn: functions that maps a GoState to a move coord (int)
'''
moves = []
prev_state, prev_action = None, None
curr_state = GoState(pachi_py.CreateBoard(board_size), BLACK)
while not curr_state.board.is_terminal:
a = (black_policy_fn if curr_state.color == BLACK else white_policy_fn)(curr_state, prev_state, prev_action)
next_state = curr_state.act(a)
moves.append((curr_state, a, next_state))
prev_state, prev_action = curr_state, a
curr_state = next_state
return moves
def _reset(self):
# Don't touch moves_log!
self.state = CustomGoState(pachi_py.CreateBoard(self.board_size),
pachi_py.BLACK)
# (re-initialize) the opponent
# necessary because a pachi engine is attached to a game via internal data in a board
# so with a fresh game, we need a fresh engine
self._reset_opponent(self.state.board)
# Let the opponent play if it's not the agent's turn
opponent_resigned = False
if self.state.color != self.player_color:
self.state, opponent_resigned = self._exec_opponent_play(
self.state, None, None)
# We should be back to the agent color
assert self.state.color == self.player_color
self.done = self.state.board.is_terminal or opponent_resigned
return self.state.board.encode()
def get(self):
"""
Get new move from PACHI given board matrix and stone color.
---
tags:
- move
consumes:
- application/json
produces:
- application/json
parameters:
- in: body
name: body
schema:
properties:
board_format:
description: format of the board, matrix or ij_history
board:
description: Board data in two dimentional array
type: array
items:
type: array
items: integer
board_size:
description: board size
stone_color:
description: Stone color for the generated move. 2:Black or 1:White
type: integer
return_board:
description: Return the board in matrix. 0:False or 1:True.
type: integer
example:
|-
{"stone_color": 1,"board": [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0,0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0,0,0,0,0,0,0,0,0,0,0,0,0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],"return_board": 1}
"""
try:
req_data = request.get_json(force=True)
print(req_data)
except:
return make_response(jsonify({"error":"Invalid payload"}), 500)
board_format = req_data.get('board_format', 'matrix')
board = req_data.get('board',[[]])
stone_color = req_data.get('stone_color',0)
return_board = req_data.get('return_board',False)
pachi_board = pachi_py.CreateBoard(19)
if board_format == 'matrix':
self.make_board_matrix(pachi_board, board)
elif board_format == 'ij_history':
stone_color = 2 if stone_color=='Black' else 1
self.make_board_ij_history(pachi_board, board)
else:
return make_response(jsonify({"error":"Invalid payload"}), 500)
_pe=pachi_py.PyPachiEngine(pachi_board,'uct','threads=8,playout=light,maximize_score')
next_move = _pe.genmove(stone_color,'')
_pe.curr_board.play_inplace(next_move, stone_color)
print(_pe.curr_board.coord_to_ij(next_move))
resp = jsonify({
'move_str':_pe.curr_board.coord_to_str(next_move),
'move_ij':_pe.curr_board.coord_to_ij(next_move),
'stone_color':stone_color,
'board':_pe.curr_board.encode()[0].tolist() if return_board else None})
del _pe
return make_response(resp,200)
from gym import envs
PATH = "./model.pt"
go_env = gym.make('gym_go:go-v0', size=5, reward_method='heuristic')
state = go_env.reset() # Reset environment and record the starting state
# Train 5x5 model
# Policy_grad(100, 10, 0.001, 0.99, go_env, False, True)
# Load trained model
policy = torch.load(PATH)
b = pachi_py.CreateBoard(5)
engine = pachi_py.PyPachiEngine(b, b'uct', b'')
# pachi_move = engine.genmove(pachi_py.WHITE, b'0')
# b.play_inplace(pachi_move, pachi_py.WHITE)
# b.play_inplace(9, pachi_py.BLACK)
# engine.notify(9, pachi_py.BLACK)
#
# pachi_move = engine.genmove(pachi_py.WHITE, b'2')
# print(pachi_move)
# # # print(b.coord_to_ij(8))
# b.play_inplace(9, pachi_py.BLACK)
# b.play_inplace(pachi_move, pachi_py.WHITE)
done = False
def test_board_sizes():
for s in [9, 19]:
b = pachi_py.CreateBoard(s)
for player in [pachi_py.BLACK, pachi_py.WHITE]:
assert len(b.get_legal_coords(player)) == s*s + 1, \
'Starting board should have size**2 + 1 legal moves'
