How to chain/compose layers in keras 2 functional API without specifying input (or input shape)

user_729288 0
Viewed 1459

I would like be able to several layers together, but before specifying the input, something like the following:

# conv is just a layer, no application
conv = Conv2D(64, (3,3), activation='relu', padding='same', name='conv')
# this doesn't work:
bn = BatchNormalization()(conv)

Note that I don't want to specify the input nor its shape if it can be avoided, I want to use this as a shared layer for multiple inputs at a later point.

Is there a way to do that? The above gives the following error:

>>> conv = Conv2D(64, (3,3), activation='relu', padding='same', name='conv')
>>> bn = BatchNormalization()(conv)
Traceback (most recent call last):
  File "/home/mitchus/anaconda3/envs/tf/lib/python3.6/site-packages/keras/engine/topology.py", line 419, in assert_input_compatibility
    K.is_keras_tensor(x)
  File "/home/mitchus/anaconda3/envs/tf/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 393, in is_keras_tensor
    raise ValueError('Unexpectedly found an instance of type `' + str(type(x)) + '`. '
ValueError: Unexpectedly found an instance of type `<class 'keras.layers.convolutional.Conv2D'>`. Expected a symbolic tensor instance.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/mitchus/anaconda3/envs/tf/lib/python3.6/site-packages/keras/engine/topology.py", line 552, in __call__
    self.assert_input_compatibility(inputs)
  File "/home/mitchus/anaconda3/envs/tf/lib/python3.6/site-packages/keras/engine/topology.py", line 425, in assert_input_compatibility
    str(inputs) + '. All inputs to the layer '
ValueError: Layer batch_normalization_4 was called with an input that isn't a symbolic tensor. Received type: <class 'keras.layers.convolutional.Conv2D'>. Full input: [<keras.layers.convolutional.Conv2D object at 0x7f3f6e54b748>]. All inputs to the layer should be tensors.

Grabbing the output of the conv layer doesn't do the trick either:

>>> bn = BatchNormalization()(conv.output)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/mitchus/anaconda3/envs/tf/lib/python3.6/site-packages/keras/engine/topology.py", line 941, in output
    ' has no inbound nodes.')
AttributeError: Layer conv has no inbound nodes.
keras-2 keras-layer keras python-3.x
3 Answers

What about using a Lambda layer.

import functools
from typing import List

from tensorflow import keras


def compose_layers(layers: List[keras.layers.Layer], **kargs) -> keras.layers.Layer:
  return keras.layers.Lambda(
    lambda x: functools.reduce(lambda tensor, layer: layer(tensor), layers, x),
    **kargs,
  )

then you can just call the compose_layers method to get the composition.

layers = [
  Conv2D(64, (3,3), activation='relu', padding='same', name='conv'),
  BatchNormalization()
]

composed_layers = compose_layers(layers, name='composed_layers')
user_40325660

You can also treat tk.Sequential as layer

import tensorflow.keras as tk
import tensorflow as tf

_layer1 = tk.layers.Conv2D(
    64, (3,3), activation='relu', 
    padding='same', name='conv'
)
_layer2 = tk.layers.BatchNormalization()

_composed_layer = tk.Sequential(
    [_layer1, _layer2]
)

_some_input = tf.random.normal((100,20,33,2))
_out = _composed_layer(_some_input)
user_56810830
Related