model.inputs and model.outputs are None when creating a sub-classed model

[noamgot]

System information

• Have I written custom code (as opposed to using a stock example script
  provided in TensorFlow): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux 4.19.112+ x86_64 - running in Google Colab
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue
  happens on a mobile device: N/A
• TensorFlow installed from (source or binary): binary (pip)
• TensorFlow version (use command below): v2.3.0-0-gb36436b087 2.3.0
• Python version: 3.6.9
• Bazel version (if compiling from source): N/A
• GCC/Compiler version (if compiling from source): N/A
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A

Describe the problem

I created a model using the sub-classing method. As I wanted to examine the model inputs and outputs (using model.inputs and model.outputs), I found them to be None. On the contrary, when I created the same model using the Sequential API, I do manage to see the inputs and outputs (using the aforementioned inputs/outputs attributes).
I also checked these attributes after predicting with the model, and calling to model.compile - nothing works, the outputs are still None.

I'd like to know how (or if?) it's possible to examine a sub-classed model's inputs and outputs this way.

Source code / logs

You can play with the code in this Colab notebook:
https://colab.research.google.com/drive/1Oq0oF9aITXNGI9iI3mTsJkmDBoKacDB8?usp=sharing

The model creating code is taken from the official TF 2 quickstart guides (with some minor modifications):
https://www.tensorflow.org/tutorials/quickstart/beginner
https://www.tensorflow.org/tutorials/quickstart/advanced

example 1 - Sequential API:

import TensorFlow as tf

model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28, 28, 1)),
  tf.keras.layers.Dense(128, activation='relu'),
  tf.keras.layers.Dense(10)
])

print(model.inputs)
print(model.outputs)

The output:

[<tf.Tensor 'flatten_5_input:0' shape=(None, 28, 28, 1) dtype=float32>]
[<tf.Tensor 'dense_11/BiasAdd:0' shape=(None, 10) dtype=float32>]

example 2 - Sub-Classed Model:

from tensorflow.keras.layers import Dense, Flatten, Conv2D
from tensorflow.keras import Model

class MyModel(Model):
  def __init__(self):
    super(MyModel, self).__init__()
    self.flatten = Flatten(input_shape=(28, 28, 1))
    self.d1 = Dense(128, activation='relu')
    self.d2 = Dense(10)

  def call(self, x):
    x = self.flatten(x)
    x = self.d1(x)
    return self.d2(x)

# Create an instance of the model
model = MyModel()
print(model.inputs)
print(model.outputs)

The output:

None
None

Note:
I noticed that when I run this code locally, I get [] instead of None, but the principle still exists.

[ravikyram]

I have tried in colab with TF version 2.3,nightly version(2.5.0-dev20201126) and was able to reproduce the issue. Please, find the gist here. Thanks!

[saikumarchalla]

Issue still exists in TF-Nightly 2.6 as well. Please find the gist here.

[kaixih]

Came across the same issue. Any suggested means to access the input list of a given subclassing model?

[falk-stefan]

The issue seems still to be a thing (2.9.1)?

[jbischof]

Sorry this bug hasn't gotten attention earlier but subclassed Models behave differently than functional or sequential ones. They are more like a keras.Layer in that their input/output signature is not determined until they are built.

However, if you want explicit inputs and outputs for a subclassed model, there is a workaround! You can declare a functional model in the __init__ and skip the call override:

from tensorflow import keras

class MyModel(keras.Model):
  def __init__(self):
    input = keras.Input(shape=(28, 28, 1), dtype="float32")
    x = keras.layers.Flatten()(input)
    x = keras.layers.Dense(128, activation='relu')(x)
    x = keras.layers.Dense(10)(x)
    super(MyModel, self).__init__(inputs=input, outputs=x)

Now the inputs/outputs will be defined:

# Create an instance of the model
model = MyModel()
print(model.inputs)
print(model.outputs)

>>> [<KerasTensor: shape=(None, 28, 28, 1) dtype=float32 (created by layer 'input_6')>]
    [<KerasTensor: shape=(None, 10) dtype=float32 (created by layer 'dense_16')>]

We use this style in keras-nlp even for complicated models like BERT.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No