CODEWITHSUNDEEP
pythontensorflowkerastensorflow2.0
desa
desa

Reputation: 1390

AttributeError: Layer has no inbound nodes - occurred in TF >= 2.4 but works in TF < 2.4?

Here there is a reduced example of the code:

class MyCustomLayer(keras.layers.Layer):
    def __init__(self, num_filters=64, kernel_size=3):
        keras.layers.Layer.__init__(self)
        self.conv_1 = keras.layers.Conv2D(filters=num_filters,
                                          kernel_size=kernel_size)
        
    def call(self, inputs):
        return self.conv_1(inputs)

x = keras.Input(shape=(None, None, 3))
my_custom_layer = MyCustomLayer()
y = my_custom_layer(x)

The following line works:

my_custom_layer.output
# Out: <KerasTensor: shape=(None, None, None, 64) dtype=float32 (created by layer 'my_custom_layer')>

However, this breaks in Tensorflow 2.4.1:

my_custom_layer.conv_1.output

with the following error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-6-74d9fc3b4fbb> in <module>
----> 1 my_custom_layer.conv_1.output

~/anaconda3/lib/python3.8/site-packages/tensorflow/python/keras/engine/base_layer.py in output(self)
   2152     """
   2153     if not self._inbound_nodes:
-> 2154       raise AttributeError('Layer ' + self.name + ' has no inbound nodes.')
   2155     return self._get_node_attribute_at_index(0, 'output_tensors', 'output')
   2156 

AttributeError: Layer conv2d has no inbound nodes.

If I run the same code in Tensorflow 2.3.1 it works without any problem. I looked through the changelog, but I did not spot what could break it. What could be the reason?

GitHub Issue #48196: https://github.com/tensorflow/tensorflow/issues/48196

Upvotes: 1

Views: 1514

Answers (2)

Chris
Chris

Reputation: 3

Here is a temporary fix to make it work (only tested on TF2.4), you can disable the use of KerasTensor with the following command:

from tensorflow.python.keras.engine import keras_tensor
keras_tensor.disable_keras_tensors()

# .... Rest of your code

Using your example, you can have access to inner component:

my_custom_layer.conv_1.output 
# <tf.Tensor 'my_custom_layer_2/conv2d_2/BiasAdd:0' shape=(None, None, None, 64) dtype=float32>

Upvotes: 0

Innat
Innat

Reputation: 17239

I've also faced similar issue. Now I've also tested your code with tf 2.0, 2.1, 2.2, 2.3 in both CPU and GPU mode and it's working. But it breaks on tf 2.4 and also tf-nightly 2.6.0-dev20210330. It seems like a bug. With tf < 2.4, the my_custom_layer.conv_1.inbound_nodes returns

[<tensorflow.python.keras.engine.node.Node at 0x7f751f301bd0>]

unlike empty [] in tf > = 2.4.

Upvotes: 3

Related Questions