我正在尝试在 2048 游戏中训练一个强化代理。我自己设计了 Env。该错误表示输入必须为 (1, 16) 形状，但数组以 (1, 4) 形状传递。我无法弄清楚在我的代码中哪里传递了具有这种形状的数组：
class CardGameEnv(Env):
def __init__(self):
self.action_space = Discrete(3)
self.observation_space = Box(low=np.array([0 for _ in range(16)]), high=np.array([np.inf for _ in range(16)]))
self._state = [0 for _ in range(16)]
self._episode_ended = False
self._score = 0
def action_spec(self):
return self.action_space
def observation_spec(self):
return self.observation_space
def reset(self):
self._state = [0 for _ in range(16)]
self._episode_ended = False
restart()
print(np.array([self._state]))
return ts.restart(np.array([self._state], dtype=np.int32))
def step(self, action):
if self._episode_ended:
# The last action ended the episode. Ignore the current action and start
# a new episode.
return self.reset()
old_score = self._score
print(action)
make_move(action)
if check_game_over():
self._episode_ended = True
self._score = get_score()
self._state = list(get_board().flatten())
score = get_score()
if self._episode_ended:
reward = - 10
return self._state, reward, self._episode_ended, {}
elif score == old_score:
reward = -10
return self._state, reward, self._episode_ended, {}
else:
reward = self._score - old_score
return self._state, reward, self._episode_ended, {}
def build_model(states, actions):
model = Sequential()
model.add(Flatten(input_shape=states))
model.add(Dense(24, activation='relu'))
model.add(Dense(24, activation='relu'))
model.add(Dense(actions, activation='softmax'))
return model
def build_agent(model, actions):
policy = BoltzmannQPolicy()
memory = SequentialMemory(limit=50000, window_length=1)
dqn = DQNAgent(model=model, memory=memory, policy=policy,
nb_actions=actions, nb_steps_warmup=10, target_model_update=1e-2)
return dqn
env = CardGameEnv()
model = build_model((1, 16), 4)
dqn = build_agent(model, 4)
dqn.compile(Adam(lr=1e-3), metrics=['mae'])
dqn.fit(env, nb_steps=50000, visualize=False, verbose=1)
这是我代码的重要部分。根据错误，我传递了一个形状为 (1, 4) 的数组，但是我到处都有一个带有 (1, 16) 的数组用于 state。
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-65-edd0afcc8057> in <module>
8 dqn = build_agent(model, 4)
9 dqn.compile(Adam(lr=1e-3), metrics=['mae'])
---> 10 dqn.fit(env, nb_steps=50000, visualize=False, verbose=1)
~\anaconda3\lib\site-packages\rl\core.py in fit(self, env, nb_steps, action_repetition, callbacks, verbose, visualize, nb_max_start_steps, start_step_policy, log_interval, nb_max_episode_steps)
166 # This is were all of the work happens. We first perceive and compute the action
167 # (forward step) and then use the reward to improve (backward step).
--> 168 action = self.forward(observation)
169 if self.processor is not None:
170 action = self.processor.process_action(action)
~\anaconda3\lib\site-packages\rl\agents\dqn.py in forward(self, observation)
222 # Select an action.
223 state = self.memory.get_recent_state(observation)
--> 224 q_values = self.compute_q_values(state)
225 if self.training:
226 action = self.policy.select_action(q_values=q_values)
~\anaconda3\lib\site-packages\rl\agents\dqn.py in compute_q_values(self, state)
66
67 def compute_q_values(self, state):
---> 68 q_values = self.compute_batch_q_values([state]).flatten()
69 assert q_values.shape == (self.nb_actions,)
70 return q_values
~\anaconda3\lib\site-packages\rl\agents\dqn.py in compute_batch_q_values(self, state_batch)
61 def compute_batch_q_values(self, state_batch):
62 batch = self.process_state_batch(state_batch)
---> 63 q_values = self.model.predict_on_batch(batch)
64 assert q_values.shape == (len(state_batch), self.nb_actions)
65 return q_values
~\anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_v1.py in predict_on_batch(self, x)
1203 ' tf.distribute.Strategy.')
1204 # Validate and standardize user data.
-> 1205 inputs, _, _ = self._standardize_user_data(
1206 x, extract_tensors_from_dataset=True)
1207 # If `self._distribution_strategy` is True, then we are in a replica context
~\anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_v1.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split, shuffle, extract_tensors_from_dataset)
2345 return [], [], None
2346
-> 2347 return self._standardize_tensors(
2348 x, y, sample_weight,
2349 run_eagerly=run_eagerly,
~\anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_v1.py in _standardize_tensors(self, x, y, sample_weight, run_eagerly, dict_inputs, is_dataset, class_weight, batch_size)
2373 if not isinstance(x, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)):
2374 # TODO(fchollet): run static checks with dataset output shape(s).
-> 2375 x = training_utils_v1.standardize_input_data(
2376 x,
2377 feed_input_names,
~\anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_utils_v1.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
661 for dim, ref_dim in zip(data_shape, shape):
662 if ref_dim != dim and ref_dim is not None and dim is not None:
--> 663 raise ValueError('Error when checking ' + exception_prefix +
664 ': expected ' + names[i] + ' to have shape ' +
665 str(shape) + ' but got array with shape ' +
ValueError: Error when checking input: expected flatten_4_input to have shape (1, 16) but got array with shape (1, 4)