CODEWITHSUNDEEP
pythontensorflowkeras
Lau
Lau

Reputation: 1465

How to assign a tf.keras.layers.layer to a class without initializing it?

I work on a class to create all sorts of symmetric AE's. I now port this class to TF 2.0 and it is more complicated than I thought. However, I use subclassing of layers and models to achieve this. Therefore I want to group several keras layers to one keras layer. But if I want to write something like this:

def __init__(self, name, keras_layer, **kwargs):
    self.keras_layer = tf.keras.layer.Conv2D
    super(CoderLayer, self).__init__(name=name, **kwargs)

I get the following error, because tf wants to use this uninitialized layer:

TypeError: _method_wrapper() missing 1 required positional argument: 'self'

I also tried to wrap this in a list, but it did not work either.

EDIT

Here is a working minimal example and the full traceback:

import tensorflow as tf
print(tf.__version__) # 2.0.0-alpha0

class CoderLayer(tf.keras.layers.Layer):

    def __init__(self, name, keras_layer):
        self.keras_layer = keras_layer
        self.out = keras_layer(12, [3, 3])
        super(CoderLayer, self).__init__(name=name)

    def call(self, inputs):
        return self.out(inputs)

inputs = tf.keras.Input(shape=(200, 200, 3), batch_size=12)
layer = CoderLayer("minimal_example", tf.keras.layers.Conv2D)

layer(inputs)

Traceback:

Traceback (most recent call last):
  File "..\baseline_cae.py", line 24, in <module>
    layer(inputs)
  File "..\AppData\Local\Continuum\anaconda3\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 581, in __call__
    self._clear_losses()
  File "..\AppData\Local\Continuum\anaconda3\lib\site-packages\tensorflow\python\training\tracking\base.py", line 456, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "..\AppData\Local\Continuum\anaconda3\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 818, in _clear_losses
    layer._clear_losses()
TypeError: _method_wrapper() missing 1 required positional argument: 'self'

Upvotes: 1

Views: 484

Answers (1)

Vlad
Vlad

Reputation: 8595

The problem is with setting not instantiated class as an attribute in a subclass of tf.keras.layers.Layer. If you remove following line

self.keras_layer = keras_layer

the code would work:

import tensorflow as tf

class CoderLayer(tf.keras.layers.Layer):
    def __init__(self, name, keras_layer):
        super(CoderLayer, self).__init__(name=name)
        self.out = keras_layer(12, [3, 3])

    def call(self, inputs):
        return self.out(inputs)

inputs = tf.keras.Input(shape=(200, 200, 3), batch_size=12)
layer = CoderLayer("minimal_example", tf.keras.layers.Conv2D)
print(layer(inputs))
# Tensor("minimal_example_3/conv2d_12/BiasAdd:0", shape=(12, 198, 198, 12), dtype=float32)

It is probably a bug. This is a similar issue that had been raised (if you put your not instantiated class into list and try to __setattr__() you will get the same exception).

This could be possible workaround if you want to use multiple layers:

class CoderLayer(tf.keras.layers.Layer):
    def __init__(self, name, layername):
        super(CoderLayer, self).__init__(name=name)
        self.layer = layername
        self.out = tf.keras.layers.__dict__[layername](1, 2)

    def call(self, inputs):
        return self.out(inputs)

inputs = tf.random.normal([1, 3, 3, 1])
layer = CoderLayer("mylayer", 'Conv2D')
layer(inputs).numpy()

Upvotes: 1

Related Questions