def read_data(data_path, max_games=None):
game_files = list(sorted(list_files(data_path, '.json')))
augmentations = list(Augmentation.iter_augmentations())
if max_games is not None:
game_files = list(game_files)[-max_games:]
print('Using game files from %s to %s' % (game_files[0], game_files[-1]))
x = []
y_policy = []
y_reward = []
for game_path in tqdm(game_files):
with open(game_path, 'r') as fin:
game_data = json.load(fin)
winner, starter, actions, policies = AlphaConnectSerializer.deserialize(game_data)
state = State.empty()
states = [state]
for action in actions:
state = state.take_action(action)
states.append(state)
states, final_state = states[:-1], states[-1]
n_samples = min(len(states), len(augmentations))
game_samples = sample(list(range(len(states))), n_samples)
for augmentation, i in zip(augmentations, game_samples):
augmentend_action_order = sorted(Action.iter_actions(), key=lambda a: a.augment(augmentation).to_int())
x.append(states[i].to_numpy(augmentation))
y_policy.append([policies[i].get(action, 0.0) for action in augmentend_action_order])
y_reward.append(winner_value(final_state.winner, states[i]))
return np.array(x), np.array(y_policy), np.array(y_reward)
def simulate(self, node: 'AlphaConnectNode', callback):
if node.state.is_end_of_game():
state_value = self.evaluate_final_state(node)
callback(state_value, None)
else:
self.queue.append((node, callback))
if len(self.queue) >= self.batch_size:
nodes, callbacks = zip(*self.queue)
array = np.concatenate(
list(map(lambda node: node.state.to_numpy(batch=True), nodes)))
pred_actions, pred_value = self.model.predict(array)
for i, callback in enumerate(callbacks):
state_value = pred_value[i].item()
action_probs = dict(zip(Action.iter_actions(),
pred_actions[i]))
callback(state_value, action_probs)
self.queue = []
def decide(self, state: State):
while True:
print('Possible actions:')
print(
format_in_action_grid(
{action: str(action)
for action in Action.iter_actions()},
cell_format='{:.2s}',
default_value=' '))
user_input = input('Choose your action: ')
try:
action = Action.from_hex(user_input)
if action in state.allowed_actions:
print()
return action
else:
print('Action %s not allowed' % action)
except ValueError:
print('User input is not an action')