def pick_model_move(self, color):
# If asked to make a move for enemy player, switch perspective as if we are them
if channel_from_color(color) == gvg.player_channel:
self.board = go_board.switch_player_perspec(self.board)
enc_board = go_board.get_encoded_board(self.board)
pred = np.asarray(self.model.predict(enc_board. \
reshape(1, gvg.board_size, gvg.board_size, gvg.enc_board_channels))). \
reshape(gvg.board_size * gvg.board_size)
prob_board = np.array(
self.model.predict(
go_board.get_encoded_board(
self.board.reshape(
gvg.board_size, gvg.board_size,
gvg.board_channels)).reshape(1, 19, 19, 20))).reshape(
(gvg.board_size, gvg.board_size))
self.last_move_probs = prob_board
move = go_learn.predict_move(self.board,
self.model,
prob_board=prob_board)
return Move(move[0], move[1])
def predict_move(orig_board, model, level=0, prob_board=None):
if prob_board is None:
prob_board = np.array(
model.predict(go_board.get_encoded_board(orig_board))).reshape(
gvg.board_size, gvg.board_size)
found_move = False
while found_move == False:
move = nanargmax(prob_board)
if go_board.make_move(orig_board,
move,
gvg.bot_channel,
gvg.player_channel,
result_only=True) == False:
prob_board[move[0]][move[1]] = -999999.0
else:
found_move = True
return move
def train_network(game_data, model):
vs = gvg.validation_split # What fraction of games are reserved for validation
hm_batches = int(len(game_data) / batch_size)
train_boards = []
train_next_moves = []
for game_index in range(len(game_data)):
board = go_board.setup_board(game_data[game_index])
for node in game_data[game_index].get_main_sequence():
board = go_board.switch_player_perspec(
board
) # Changes player perspective, black becomes white and vice versa
node_move = node.get_move()[1]
if node_move is not None:
train_boards.append(go_board.get_encoded_board(board))
next_move = np.zeros(gvg.board_size * gvg.board_size).reshape(
gvg.board_size, gvg.board_size)
next_move[node_move[0], node_move[
1]] = gvg.filled # y = an array in the form [board_x_position, board_y_position]
train_next_moves.append(
next_move.reshape(gvg.board_size * gvg.board_size))
board = go_board.make_move(
board, node_move, gvg.bot_channel,
gvg.player_channel) # Update board with new move
if board is None:
print("ERROR! Illegal move, {}, while training".format(
node_move[1]))
if len(train_boards
) >= batch_size: # Send chunk to GPU at batch limit
model.fit(
{"input": train_boards[:int(-len(train_boards) * vs)]}, {
"target":
train_next_moves[:int(-len(train_boards) * vs)]
},
validation_set=({
"input":
train_boards[int(-len(train_boards) * vs):]
}, {
"target":
train_next_moves[int(-len(train_boards) * vs):]
}),
n_epoch=1,
batch_size=gvg.train_batch_size,
snapshot_step=7500,
show_metric=True)
train_boards = []
train_next_moves = []
# Finish of what is remaining in the batch
model.fit({"input": train_boards[:int(-len(train_boards) * vs)]},
{"target": train_next_moves[:int(-len(train_boards) * vs)]},
validation_set=({
"input":
train_boards[int(-len(train_boards) * vs):]
}, {
"target":
train_next_moves[int(-len(train_boards) * vs):]
}),
n_epoch=1,
batch_size=gvg.train_batch_size,
snapshot_step=7500,
show_metric=True)
print("Done loading {} games".format(len(games)))
print("Being data processing...")
train_boards = []
train_next_moves = []
for game_index in range(len(games)):
board = go_board.setup_board(games[game_index])
for node in games[game_index].get_main_sequence():
board = go_board.switch_player_perspec(
board
) # Changes player perspective, black becomes white and vice versa
node_move = node.get_move()[1]
if node_move is not None:
train_boards.append(go_board.get_encoded_board(board))
next_move = np.zeros(gvg.board_size * gvg.board_size).reshape(
gvg.board_size, gvg.board_size)
next_move[node_move[0], node_move[
1]] = gvg.filled # y = an array in the form [board_x_position, board_y_position]
train_next_moves.append(
next_move.reshape(gvg.board_size * gvg.board_size))
board = go_board.make_move(
board, node_move, gvg.bot_channel,
gvg.player_channel) # Update board with new move
if board is None:
print("ERROR! Illegal move, {}, while training".format(
node_move))
print("Finished data processing...")