"No common supertype of ..." while loading model

[ahasselbring]

System information.

• Have I written custom code (as opposed to using a stock example script provided in Keras): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Mint 21
• TensorFlow installed from (source or binary): binary (tensorflow-cpu via pipenv)
• TensorFlow version (use command below): v2.9.1-132-g18960c44ad3 2.9.2
• Python version: 3.10
• Bazel version (if compiling from source): -
• GPU model and memory: -
• Exact command to reproduce: Given the standalone code below saved as minimal-example.py, run python minimal-example.py

Describe the problem.

I can't load a model when passing a variable to a submodel. See the example below. This explicitly tells me to "[...] file a bug if [...] being hindered by this error." Code like that worked with previous versions of TensorFlow (2.7 I think).

Describe the current behavior.

Loading fails with No common supertype of TensorSpec(...) and VariableSpec(...).

Describe the expected behavior.

The model is loaded without errors.

Contributing.

• Do you want to contribute a PR? (yes/no): no

Standalone code to reproduce the issue.

import tensorflow as tf

class SubModel(tf.keras.models.Model):
    def call(self, x):
        return ()  # My actual model returns something depending on x here, but even this fails.

class MainModel(tf.keras.models.Model):
    def __init__(self):
        super().__init__()
        self.sub_model = SubModel()
        self.x = tf.Variable(tf.zeros((1,)))

    def call(self, _):
        return self.sub_model(self.x)

model = MainModel()
model(())
model.save("model")
model = tf.keras.models.load_model("model")

Source code / logs.

Output of the above script:

Traceback (most recent call last):
  File ".../minimal-example.py", line 19, in <module>
    model = tf.keras.models.load_model("model")
  File ".../.venv/lib/python3.10/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File ".../.venv/lib/python3.10/site-packages/keras/saving/saved_model/load.py", line 1151, in common_spec
    raise TypeError(f'No common supertype of {x} and {y}.')
TypeError: No common supertype of TensorSpec(shape=(None,), dtype=tf.float32, name=None) and VariableSpec(shape=(1,), dtype=<dtype: 'float32'>, trainable=True).

[sushreebarsa]

@ahasselbring I tried to replicate the issue on colab and faced following error;

ValueError: Types are not compatible: TensorSpec(shape=(None,), dtype=tf.float32, name=None) with type of <class 'tensorflow.python.framework.tensor_spec.TensorSpec'> vs VariableSpec(shape=(1,), dtype=tf.float32, name='x') with type of <class 'tensorflow.python.ops.resource_variable_ops.VariableSpec'>.

Could you find the gist here and confirm the same? Thank you!

[ahasselbring]

Indeed I get that error when I use TensorFlow 2.8, but with both 2.9 and 2.10 the error is as in the issue description:

No common supertype of TensorSpec(shape=(None,), dtype=tf.float32, name=None) and VariableSpec(shape=(1,), dtype=<dtype: 'float32'>, trainable=True).

coming from here (2.9) or here (2.10). Unfortunately I can't test with 2.7 anymore (I'm quite sure things like that worked with that version - maybe not this minimal example, but generally, passing variables to submodels, letting them do calculations and return them).

[google-ml-butler[bot]]

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

[gowthamkpr]

@ahasselbring With tensorflow 2.7, there is a warning that states WARNING:tensorflow:No training configuration found in save file, so the model was *not* compiled. Compile it manually.

You can find the gist here.

[ahasselbring]

That warning can easily removed by adding compile=False to the load_model call in 2.7 (or by actually doing some training with a useful model, but that doesn't make sense in a minimal example). In newer versions, the error occurs regardless of that option.

[haifeng-jin]

Summry: This model cannot be save and load. (Minor modification made to the original snippet, tested with colab)

class SubModel(tf.keras.models.Model):
    def call(self, x):
        return x

class MainModel(tf.keras.models.Model):
    def __init__(self):
        super().__init__()
        self.sub_model = SubModel()
        self.x = tf.Variable(tf.zeros((1,)))

    def call(self, x):
        self.sub_model(self.x)
        return x

[rchao]

Thanks for reporting the issue - at this point we're lacking bandwidth to look further into this and community contribution for a potential fix is welcomed.

[ahasselbring]

I am working around this now by passing variable + tf.zeros_like(variable) instead of variable.

[georgeyw]

I have a similar issue with a model that uses RaggedTensors. I see an error message that looks something like "No common supertype of TensorSpec and RaggedTensorSpec".

I seem to be able to fix this by rolling back the code for infer_inputs_from_restored_call_function from r2.10 (found here: https://github.com/keras-team/keras/blob/r2.10/keras/saving/saved_model/load.py#L1293) to r2.7 (found here: https://github.com/keras-team/keras/blob/r2.7/keras/saving/saved_model/load.py#L1172, which also requires adding back get_common_shape right above it), but leaving everything else in r2.10 the same. It looks to me like this is the only place where this function is used.

I didn't seem to notice any problems with this in the local testing that I did, does anyone have context for why this was changed or know of side effects of rolling it back?

[ThibaultDef]

By looking directly into the code, I saw that the error comes from the function most_specific_common_supertype (within https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/framework/type_spec.py). If I understood well, I compares types between tensors. If a type is not matching between two tensors (for example one is of type float32 and another one is of type float16), then the error raises.