Adding TensorFlow Hub KerasLayer to Sequential Model Raises ValueError

[ruddyscent]

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

2.16.1

Custom code

No

OS platform and distribution

Ubuntu 22.04.3 LTS

Mobile device

No response

Python version

Python 3.11.0rc1

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

12.3.0

GPU model and memory

No response

Current behavior?

I can execute the following code without any issues in TensorFlow 2.15.0 and TensorFlow Hub 1.16.1. However, when I upgrade the TensorFlow version to 2.16.0 or above, I encounter an error stating that KerasLayer cannot be added to the Sequential model.

Standalone code to reproduce the issue

import tensorflow as tf
import tensorflow_hub as hub

image_size = 224
URL = "https://tfhub.dev/tensorflow/efficientnet/b0/feature-vector/1"

model = tf.keras.Sequential([
        hub.KerasLayer(URL, input_shape=(image_size, image_size, 3))
])

Relevant log output

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[29], line 1
----> 1 model = tf.keras.Sequential([
      2         feature_extractor,
      3         tf.keras.layers.Dense(2, activation = 'softmax')
      4 ])
      6 model.build([None, image_size, image_size, 3])
      7 model.summary()

File /usr/local/lib/python3.10/dist-packages/keras/src/models/sequential.py:70, in Sequential.__init__(self, layers, trainable, name)
     68 if layers:
     69     for layer in layers:
---> 70         self.add(layer, rebuild=False)
     71     self._maybe_rebuild()

File /usr/local/lib/python3.10/dist-packages/keras/src/models/sequential.py:92, in Sequential.add(self, layer, rebuild)
     90         layer = origin_layer
     91 if not isinstance(layer, Layer):
---> 92     raise ValueError(
     93         "Only instances of `keras.Layer` can be "
     94         f"added to a Sequential model. Received: {layer} "
     95         f"(of type {type(layer)})"
     96     )
     97 if not self._is_layer_name_unique(layer):
...
    101         "the name of a layer in this model. Update the `name` argument "
    102         "to pass a unique name."
    103     )

ValueError: Only instances of `keras.Layer` can be added to a Sequential model. Received: <tensorflow_hub.keras_layer.KerasLayer object at 0x7a4ac7e30f40> (of type <class 'tensorflow_hub.keras_layer.KerasLayer'>)

[Aloqeely]

After investigating the code, I found a potential cause of the issue in tensorflow_hub/keras_layer.py lines 26-31:

# Use Keras 2.
version_fn = getattr(tf.keras, "version", None)
if version_fn and version_fn().startswith("3."):
  import tf_keras as keras
else:
  keras = tf.keras

Depending on the Keras version, the module might either import tf_keras or directly use tf.keras, the former causes the isinstance(layer, Layer) check in Sequential.add to return False for hub.KerasLayer, even though it inherits from keras.layers.Layer

[SuryanarayanaY]

Hi @ruddyscent ,

TFHub has dependency on tf_keras package (i.e Keras2) as per the setup.py of TFHub.

https://github.com/tensorflow/hub/blob/ff72e25fef44bd67d5c14fb5328aa44e303e3404/tensorflow_hub/pip_package/setup.py#L31

Since TF2.16 comes with Keras3 the problem arises. As a workaround, you can install tf_keras package and set environment variable TF_USE_LEGACY_KERAS=1 to ensure Keras2 will be used with tf.keras.

[SuryanarayanaY]

After investigating the code, I found a potential cause of the issue in tensorflow_hub/keras_layer.py lines 26-31:

# Use Keras 2.
version_fn = getattr(tf.keras, "version", None)
if version_fn and version_fn().startswith("3."):
  import tf_keras as keras
else:
  keras = tf.keras

Depending on the Keras version, the module might either import tf_keras or directly use tf.keras, the latter causes the isinstance(layer, Layer) check in Sequential.add to return False for hub.KerasLayer, even though it inherits from keras.layers.Layer

Hi, When I tried with TF2.16v on Colab environment the error stack seems generated from the _ensure_keras_2_importable() function from tensorflow_hub/__init__.py as per attached gist.

[SuryanarayanaY]

# Use Keras 2.
version_fn = getattr(tf.keras, "version", None)
if version_fn and version_fn().startswith("3."):
  import tf_keras as keras
else:
  keras = tf.keras

Even this code seems confusing to me. If version_fn results in Keras3 then it tries to import tf_keras to use Keras2 else it assumes Keras2 is alreday there(i.e TF<=2.15v) hence it takes keras=tf.keras.

But version_fn = getattr(tf.keras, "version", None) returns None making this to use tf.keras (i.e Keras3) which ic not compatible with TF_Hub.

[Aloqeely]

# Use Keras 2.
version_fn = getattr(tf.keras, "version", None)
if version_fn and version_fn().startswith("3."):
  import tf_keras as keras
else:
  keras = tf.keras

Even this code seems confusing to me. If version_fn results in Keras3 then it tries to import tf_keras to use Keras2 else it assumes Keras2 is alreday there(i.e TF<=2.15v) hence it takes keras=tf.keras.

But version_fn = getattr(tf.keras, "version", None) returns None making this to use tf.keras (i.e Keras3) which ic not compatible with TF_Hub.

Actually, version_fn is not None, that code works as intended, which is to use Keras 2

tf.keras.Sequential is on Keras 3. and hub.KerasLayer is on Keras 2, so I believe due to the version difference, it causes isinstance(layer, Layer) to fail and return False

[SuryanarayanaY]

# Use Keras 2.
version_fn = getattr(tf.keras, "version", None)
if version_fn and version_fn().startswith("3."):
  import tf_keras as keras

From the code above if version_fn is not None and if it starts with "3" (which means Keras 3 found) in that case it is importing tf_keras package as keras. Please note that tf_keras package is for Keras2. This indicates TFHub supports only Keras2 package.

The else part it assumes that version_fn doesn't starts with "3" which case it assumes Keras2 installed with TF package and it marks keras=tf.keras. But for any case if version_fn becomes None and if Keras3 installed with TF then tf.keras will become Keras3 which might a problem.

@Aloqeely , Could you please import TF2.16 and confirm what will be the version_fn output ?

[Aloqeely]

Here you go

import tensorflow as tf

version_fn = getattr(tf.keras, "version", None)
print("TF Version: " + tf.__version__)
print("TF Keras Version: " + version_fn())

Output:

TF Version: 2.16.1
TF Keras Version: 3.0.5