Question

我正在为 OpenAI lunarLander-v2 环境训练模型。我已经使用 Sequential 模型成功地做到了这一点，但是在尝试使用功能模型时，我遇到了一些与 tensorshapes 不兼容的错误。这是 Agent 类的代码，我认为这个问题与 done_list 和 next_states 的形状不兼容，但我不确定如何重塑这些张量以使其工作。

class DQAgent(Agent):
def __init__(self, env, config):
    Agent.__init__(self, env, config)
    
    self.memory = deque(maxlen=self.config.memory_size)
    self.model = self.initialize()

def initialize(self):
    
    inputs = Input(shape=(8,))
    
    dense = Dense(self.config.layer_size * self.config.input_layer_mult, activation = relu)
    x = dense(inputs)
    x = Dense(self.config.layer_size, activation = relu)(x)
    
    outputs = layers.Dense(self.action_space_size, activation = linear)(x)
    
    model = keras.Model(inputs = inputs, outputs = outputs, name = self.name)

    model.compile(loss = mean_squared_error, optimizer = Adam(lr = self.config.learning_rate))
    model.summary()

    return model

def policyAct(self, state):
    predicted_actions = self.model.predict(state)
    return np.argmax(predicted_actions[0])

def addToMemory(self, state, action, reward, next_state, done):
    self.memory.append((self, state, action, reward, next_state, done))
    
def sampleFromMemory(self):
    sample = np.random.sample(self.memory, self.config.batch_size)
    return sample

def extractFromSample(self, sample):
    states = np.array([i[0] for i in sample])
    actions = np.array([i[1] for i in sample])
    rewards = np.array([i[2] for i in sample])
    next_states = np.array([i[3] for i in sample])
    done_list = np.array([i[4] for i in sample])
    states = np.squeeze(states)
    next_states = np.squeeze(next_states)
    
    
    return np.squeeze(states), actions, rewards, next_states, done_list
    
def updateReplayCount(self):
    self.config.replay_counter += 1
    self.config.replay_counter = self.replay_counter % self.config.replay_step_size

def learnFromMemory(self):
    if len(self.memory) < self.config.batch_size or self.config.replay_counter != 0:
        return
    if np.mean(self.training_episode_rewards[-10:]) > 100:
        return
    sample = self.sampleFromMemory()

    states, actions, rewards, next_states, done_list = self.extractFromSample(sample)
    targets = rewards + self.config.gamma * (np.amax(self.model.predict_on_batch(next_states), 
                                                     axis=1)) * (1 - (done_list))
    
    target_vec = self.model.predict_on_batch(states)
    indexes = np.array([i for i in range(self.config.batch_size)])
    target_vec[[indexes], [actions]] = targets
    self.model.fit(states, target_vec, epochs=1, verbose=0)
    
def save(self, name):
    self.model.save(name)

在使用 Sequential API 而不是函数式 API 创建模型时，类似的代码工作正常。我对此很陌生，对 SO 也是如此，非常感谢任何帮助。

<块引用>

警告：tensorflow:Model 是用形状 (None, 8) 构建的，用于输入 Tensor("input_10:0", shape=(None, 8), dtype=float32)，但它在形状不兼容的输入上被调用（无，1）。值错误：dense_72 层的输入 0 与层不兼容：输入形状的预期轴 -1 具有值 8，但收到的输入形状为 [None, 1]

来自顺序实现的模型，运行没有问题（其余代码相同）

def initialize_model(self):
    model = Sequential()
    
   
    model.add(Dense(self.config.layer_size*self.config.input_layer_mult, input_dim = self.observation_space_dim, activation=relu))
    
    
    for i in range(self.config.deep_layers):
        model.add(Dense(self.config.layer_size, activation=relu))
    
    
    model.add(Dense(self.action_space_dim, activation=linear))
    
    
    model.compile(loss=mean_squared_error, optimizer=Adam(lr=self.config.learning_rate))

    print(model.summary())
    
    return model

Answer 1

从这里开始，[https://stackoverflow.com/questions/64512293/input-dense-is-incompatible-with-the-layer-invalid-shape][1]

输入形状应该是 (1, )

ValueError：密集层的输入 0 与层不兼容：预期轴 -1 具有值 8，但收到的输入具有形状 [None, 1]

1 个答案: