Question

我遵循Denny Britz's TF implementation作为准则，使用keras从零开始实现了基本的REINFORCE算法。我正在尝试使用多代理设置，一次训练两个代理（发送者和接收者）。对于两个代理，train_on_batch的第一次调用运行都没有问题。但是，当第二次调用train_on_batch时，它会崩溃，并引发以下错误：

FailedPreconditionError:  Error while reading resource variable _AnonymousVar18 from Container: 
localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/_AnonymousVar18/class 
tensorflow::Var does not exist.
     [[node mul_3/ReadVariableOp (defined at D:\ProgramData\Anaconda3\envs\jupyter-mess\lib\site-packages\keras\backend\tensorflow_backend.py:3014) ]] 
[Op:__inference_keras_scratch_graph_1235]

Function call stack:
keras_scratch_graph

这里是Sender模型，这是一个具有自定义丢失的简单前馈网络。接收器网络非常相似。

class Sender:
    def __init__(self, n_images, input_image_shape, embedding_size, vocabulary_size, temperature):
        self.reset_memory()
        image_inputs = [layers.Input(shape=input_image_shape, dtype="float32") 
            for i in range(n_images)]
        image_embedding_layer = layers.Dense(embedding_size)
        sigmoid = layers.Activation("sigmoid")
        output_layer = layers.Dense(vocabulary_size)
        temperature_layer = layers.Lambda(lambda x: x / temperature)
        softmax = layers.Softmax()
        
        y = [image_embedding_layer(x) for x in image_inputs]
        y = [sigmoid(x) for x in y]
        y = layers.concatenate(y, axis=-1)
        y = output_layer(y)
        y = temperature_layer(y)
        y = softmax(y)
        
        self.model = Model(image_inputs, y)
        
        index = layers.Input(shape=[1], dtype="int32")
        y_selected = layers.Lambda(
            lambda probs_index: tf.gather(*probs_index, axis=-1),
            )([y, index])
        
        def loss(target, prediction):
            return - K.log(prediction) * target
        
        self.model_train = Model([*image_inputs, index], y_selected)
        self.model_train.compile(loss=loss, optimizer=OPTIMIZER)
        
    def predict(self, state):
        return self.model.predict_on_batch(x=state)

    def update(self, state, action, target):
        x = [*state, action]
        return self.model_train.train_on_batch(x=x, y=target)

这是执行：

for episode in range(1, N_EPISODES + 1):
    game.reset()
    sender_state = game.get_sender_state(n_images=N_IMAGES, unique_categories=True, expand=True)
    sender_probs = sender.predict(state=sender_state)
    sender_probs = np.squeeze(sender_probs)
    sender_action = np.random.choice(np.arange(len(sender_probs)), p=sender_probs)
    
    receiver_state = game.get_receiver_state(sender_action, expand=True)
    receiver_probs = receiver.predict(receiver_state)
    receiver_probs = np.squeeze(receiver_probs)
    receiver_action = np.random.choice(np.arange(len(receiver_probs)), p=receiver_probs)
    
    sender_reward, receiver_reward, success = game.evaluate_guess(receiver_action)

    sender.update(sender_state, np.asarray([sender_action]), np.asarray([sender_reward]))
    receiver.update(receiver_state, np.asarray([receiver_action]), np.asarray([receiver_reward]))

当我在使用 keras 2.3.1 的jupyter笔记本上本地运行该实现时，该实现会崩溃，但在使用 keras 2.4.3 的Google Colab上，该实现运行良好。

此外，当我禁用其中一个代理程序时，该程序对于另一个代理程序运行时没有错误，但是我无法以这种方式进行训练。

是否可以查看导致故障的变量？我是否需要为每个代理或类似的东西运行单独的会话？

查找导致“ FailedPreconditionError：读取资源变量时出错”的变量

0 个答案: