Keras：张量对象没有属性“_keras_history”

Question

similar 先前问题的大多数答案都建议将有问题的张量包装在 Lambda 层中。但是，我已经这样做了（并尝试了修复的许多变体），它仍然抛出相同的错误。 我当前模型定义的伪代码如下所示：

# [previous layers of model definition not shown here for simplicity]
out_duration = Reshape((30, 3))(out_duration)
out_position = Reshape((30, 3))(out_position)
low = tf.constant([(30x3) numpy array of lower bounds)]) # can't use K.clip here because the lower bound is different for every element
high = tf.constant([(30x3) numpy array of upper bounds)])
clipped_out_position = Lambda(lambda x: tf.clip_by_value(*x), output_shape=out_position.get_shape().as_list())([out_position, low, high])
model = Model(inputs=x, outputs=[out_duration, clipped_out_position]

错误输出：

   File "multitask_inverter.py", line 107, in <module>
    main()
  File "multitask_inverter.py", line 75, in main
    model = Model(inputs=x, outputs=[out_duration, clipped_out_position])
  File "/om/user/lnj/openmind_env/tensorflow-gpu/lib/python3.5/site-packages/keras/legacy/interfaces.py", line 88, in wrapper
    return func(*args, **kwargs)
  File "/om/user/lnj/openmind_env/tensorflow-gpu/lib/python3.5/site-packages/keras/engine/topology.py", line 1705, in __init__
    build_map_of_graph(x, finished_nodes, nodes_in_progress)
  File "/om/user/lnj/openmind_env/tensorflow-gpu/lib/python3.5/site-packages/keras/engine/topology.py", line 1695, in build_map_of_graph
    layer, node_index, tensor_index)
  File "/om/user/lnj/openmind_env/tensorflow-gpu/lib/python3.5/site-packages/keras/engine/topology.py", line 1665, in build_map_of_graph
    layer, node_index, tensor_index = tensor._keras_history
AttributeError: 'Tensor' object has no attribute '_keras_history'

Answer 1

您可以将low 和high 提供给具有Lambda(..., arguments={'low': low, 'high': high}) 的层。来自Lambda的文档：

arguments：要传递给的关键字参数的可选字典 函数。

例如，

clipped_out_position = Lambda(lambda x, low, high: tf.clip_by_value(x, low, high),
                              arguments={'low': low, 'high': high})(out_position)

---

编辑：

下面是测试这一层的更完整示例：

x = Input(shape=(100,))
out_duration = Dense(90)(x)
out_position = Dense(90)(x)
out_duration = Reshape((30, 3))(out_duration)
out_position = Reshape((30, 3))(out_position)

low = tf.constant(np.random.rand(30, 3).astype('float32'))
high = tf.constant(1 + np.random.rand(30, 3).astype('float32'))
clipped_out_position = Lambda(lambda x, low, high: tf.clip_by_value(x, low, high),
                              arguments={'low': low, 'high': high})(out_position)

model = Model(inputs=x, outputs=[out_duration, clipped_out_position])
model.compile(loss='mse', optimizer='adam')
model.fit(np.random.rand(5000, 100), [np.random.rand(5000, 30, 3), np.random.rand(5000, 30, 3)])