张量对象在急切执行时不可迭代......同时使用 Keras 形状函数

Question

我正在尝试执行以下操作。如果有更好的方法来实现相同的想法，我需要在执行过程中得到错误的帮助，一般来说，建议，如下所述。

网络模型 - https://i.imgur.com/LdKwXRP.png

我有可变数量的网络，例如在该图中，有 3 个图示。每一层都是按顺序执行的。

其他密集层的输出用作与当前正在执行的密集层的输出进行点积的输入。 其实现是通过将这些中间层的输出存储在一个名为 ActorNetwork.k_list 的类变量列表中来完成的。

在执行过程中，我得到了错误 -

Traceback (most recent call last):
  File "training-code.py", line 80, in <module>
    main(args)
  File "training-code.py", line 36, in main
    ActorNetwork(sess, observation_dim[i], action_dim[i], float(args['actor_lr']), float(args['tau']), n))
  File "/home/rangwala/maddpg-attn/actorcriticv2.py", line 30, in __init__
    self.mainModel, self.mainModel_weights, self.mainModel_state = self._build_model()
  File "/home/rangwala/maddpg-attn/actorcriticv2.py", line 55, in _build_model
    keys = Input(shape=(K.shape(ActorNetwork.k_list,)))
  File "/home/rangwala/anaconda3/envs/comm-nav-cpu/lib/python3.6/site-packages/keras/engine/input_layer.py", line 171, in Input
    batch_shape = (None,) + tuple(shape)
  File "/home/rangwala/anaconda3/envs/comm-nav-cpu/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 431, in __iter__
    "Tensor objects are not iterable when eager execution is not "
TypeError: Tensor objects are not iterable when eager execution is not enabled. To iterate over this tensor use tf.map_fn.

代码 -

class ActorNetwork(object):
"""
Implements actor network
"""

    k_list = [] #stores the values from the intermediate layers

# More code in between

    def _build_model(self):
        input_obs = Input(shape=(self.state_dim,))
        keys = Input(shape=(K.shape(ActorNetwork.k_list, ))) 
        #k_list is a class variable, of the ActorNetwork class.

        h = Dense(400)(input_obs)
        h = Activation('relu')(h)

        query = Dense(self.n_attn, name="keys_layer")(h)
        ky_list = ActorNetwork.k_list
        keys_list = ky_list.pop(self.n) #remove own entry from the list, for dot product
        concat_layer = Concatenate(axis=1)
        all_agents = concat_layer(keys_list)

        attn = tf.einsum('i, i->ij', [query, all_agents]) / self.temper #dot product

        attn = Activation('softmax')(attn)
        attn = Dropout(0.1)(attn)

        attn_out = tf.einsum('ik, k->i', [all_agents, attn])
        attn_add = Lambda(lambda x: x[0] + x[1])([query, attn_out]) #add own value to dot product value

        h = Dense(self.action_dim)(attn_add)
        pred = Activation('tanh')(h)
        pred = BatchNormalization()(pred)

        model = Model(inputs=[input_obs, keys], outputs=pred)
        model.compile(optimizer='Adam', loss='categorical_crossentropy')

        attn_layer_out = model.get_layer("keys_layer").output
        ActorNetwork.k_list[self.n] = attn_layer_out 

        return model, model.trainable_weights, input_obs

Answer 1

看起来 keras 正在尝试将形状张量视为可迭代的 - 它只能在急切模式下进行。假设您知道ActorNetwork.k_list 的静态形状，您可以使用k_list.shape.as_list() 将其转换为列表。

我怀疑它在keys 行，所以试试

keys = Input(shape=ActorNetwork.k_list.shape.as_list()))