Even in eager mode, Keras is passing custom models non-eager tensors in 'fit'

[malmaud]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 10.14
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): '1.14.1-dev20190301'
• Python version: 3.7

Describe the current behavior
When calling keras.model.fit on a custom model, it seems the model is passed a graph-mode tensor instead of an eager tenser, even when in eager mode.

Describe the expected behavior
If in eager mode, the tensors passed to the call method of a custom model should be eager tensors. Otherwise, the advantages of eager mode, like the ability to use native control flow, are lost.

Code to reproduce the issue

import tensorflow as tf
tf.enable_v2_behavior()
from tensorflow import keras

class MyModel(keras.Model):
    def call(self, x):
        if x > 0:
            return x + 1
        else:
            return x - 1
            
m = MyModel()
m(tf.constant(0))  # This works, returns -1 as expected
m.compile(loss='mse', optimizer='sgd')
m.fit(tf.constant(0), tf.constant(1)) # This fails

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-55-2c7beacc3d4f> in <module>
----> 1 m.fit(tf.constant(0), tf.constant(1))

/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)
    804         steps=steps_per_epoch,
    805         validation_split=validation_split,
--> 806         shuffle=shuffle)
    807 
    808     # Prepare validation data.

/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split, shuffle, extract_tensors_from_dataset)
   2501       else:
   2502         cast_inputs = x_input
-> 2503       self._set_inputs(cast_inputs)
   2504     else:
   2505       y_input = y

/anaconda3/lib/python3.7/site-packages/tensorflow/python/training/tracking/base.py in _method_wrapper(self, *args, **kwargs)
    454     self._setattr_tracking = False  # pylint: disable=protected-access
    455     try:
--> 456       result = method(self, *args, **kwargs)
    457     finally:
    458       self._setattr_tracking = previous_value  # pylint: disable=protected-access

/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _set_inputs(self, inputs, outputs, training)
   2773             outputs = self.call(inputs, training=training)
   2774           else:
-> 2775             outputs = self.call(inputs)
   2776           # Reset to the previously saved value. If `call()` had `add_metric`
   2777           # or `add_loss`, then `_contains_symbolic_tensors` will have been set

<ipython-input-52-94b7a4f52815> in call(self, x)
      1 class MyModel(keras.Model):
      2     def call(self, x):
----> 3         if x > 0:
      4             return x+1
      5         else:

/anaconda3/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in __bool__(self)
    658       `TypeError`.
    659     """
--> 660     raise TypeError("Using a `tf.Tensor` as a Python `bool` is not allowed. "
    661                     "Use `if t is not None:` instead of `if t:` to test if a "
    662                     "tensor is defined, and use TensorFlow ops such as "

TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.

[malmaud]

Note that even adding run_eagerly=True to the compile call doesn't resolve this.

[jekbradbury]

I believe you need to pass dynamic=True to the model constructor:

m = MyModel(dynamic=True)

Otherwise Keras assumes that your model is able to take symbolic tensors (and uses this for shape inference) even if you've asked it to compile in eager mode. This was added in d0373dc, which has a little more context.

[jekbradbury]

There is a bug here, though, in that Keras should probably tell you why this is failing.

[malmaud]

Thanks, that does the trick!

There also might be a bit a documentation issue here - I don't recall this being prominently highlighted in the new tutorials and overviews of eager mode/keras.

[pavithrasv]

Thank you @jekbradbury for the quick response on this! We will definitely need to throw an error explaining the issue here. I will look into that.

@malmaud I agree that this is not documented well. This concept is very new and we are working to see if we can remove the need for this hence the lack of documentation.

[wevonosky]

Can we get some more explanation for how to add dynamic=True and what that does?

[brett-daley]

I found a solution that works for me and posted it in this StackOverflow answer.

tl;dr set the run_eagerly attribute manually after compiling your model:

model.compile()
model.run_eagerly = True

[yin1999]

@brett-daley It works! Thank you.

[surGeonGG]

I am having the exact same problem with keras version 2.3.0-tf and tensorflow version 2.2.0. Adding dynamic=True gives me this error for a functional model:

TypeError: ('Functional models may only specify 'name' and 'trainable' keyword arguments during initialization. Got an unexpected argument:', 'dynamic')

[surGeonGG]

I was able to fix my issue by subclassing tf.Keras.Model. Also, there was no need to pass dynamic=True to the model constructor; run_eagerly=True when compiling works.

[talpay]

@surGeonGG Can you be more specific please? I assume you did more than just simply subclassing because propagating the init via a super-call shouldn't make any difference.

[PrattJena]

I am having the same issue when running this on 2.3.0. Doing this also doesnt resolve the issue

model.compile()
model.run_eagerly = True

[dannyng]

Same with PrattJena, running on 2.3.0 and have same issue even with run_eagerly = True

[Carterbouley]

I also have this same issue. Any pointers towards a solution?

[tensorflowbutler]

Hi There,

We are checking to see if you still need help on this, as you are using an older version of tensorflow which is officially considered end of life . We recommend that you upgrade to the latest 2.x version and let us know if the issue still persists in newer versions. Please open a new issue for any help you need against 2.x, and we will get you the right help.

This issue will be closed automatically 7 days from now. If you still need help with this issue, please provide us with more information.

[Carterbouley]

Hi Alfred, This issue has been resolved. Thanks so much for checking in though.

…

On Fri, Feb 19, 2021 at 7:10 PM Alfred Sorten Wolf ***@***.***> wrote: Hi There, We are checking to see if you still need help on this, as you are using an older version of tensorflow which is officially considered end of life . We recommend that you upgrade to the latest 2.x version and let us know if the issue still persists in newer versions. Please open a new issue for any help you need against 2.x, and we will get you the right help. This issue will be closed automatically 7 days from now. If you still need help with this issue, please provide us with more information. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#26268 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKEAU4WCOELDST4GRUNK6FDS72ZTZANCNFSM4G3H55ZQ> .

[ymodak]

Closing this issue since it is resolved. Feel free to reopen if necessary. Thanks!

[google-ml-butler]

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

[eyildiz-ugoe]

This isn't solved. I was able to reproduce the error even after passing the suggested line after the compile() function is called.