def replay(self, wps, pi_mcts, board_logs, plus_turns, weights,
batch_size: int, beta: float) -> None:
inputs = np.zeros((batch_size, 7, 5, 3))
policy_true = np.zeros((batch_size, 315))
values_true = np.zeros((batch_size))
input_weights = np.zeros((batch_size))
indices = np.random.choice(np.arange(len(wps)),
size=batch_size,
replace=False)
mini_batch = [(wps[i], pi_mcts[i], board_logs[i], plus_turns[i],
weights[i]) for i in indices]
for i, (winner, pi, board, plus_turn, weight) in enumerate(mini_batch):
gs = GameState()
gs.board = board
inputs[i] = gs.to_inputs(flip=not plus_turn) # shape=(4, 5, 5)
policy_true[i] = pi**beta
values_true[i] = winner
input_weights[i] = weight
# epochsは訓練データの反復回数、verbose=0は表示なしの設定
self.model.fit(inputs, [policy_true, values_true],
sample_weight=input_weights,
epochs=1,
verbose=0,
shuffle=True)
