Keras: Adam Optimizer no longer working with TensorFlow 1.6.0

[pierluigiferrari]

Keras version: 2.1.5
TensorFlow version: 1.6.0

Trying to use an Adam optimizer gives the following error below. This error does not occur with TensorFlow 1.4.1 with everything else being exactly the same.

I guess it has to do with the following change that is documented in the TensorFlow 1.6.0 release notes:

New Optimizer internal API for non-slot variables. Descendants of AdamOptimizer that access _beta[12]_power will need to be updated.

This is the error:

WARNING:tensorflow:Variable *= will be deprecated. Use variable.assign_mul if you want assignment to the variable value or 'x = x * y' if you want a new python Tensor object.
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-24-22d018627654> in <module>()
     21                                                              cooldown=0)],
     22                               validation_data = val_generator,
---> 23                               validation_steps = ceil(num_val_samples/batch_size))
     24 
     25 # TODO: Set the filename (without the .h5 file extension!) under which to save the model and weights.

~/Code/tensorflow/lib/python3.6/site-packages/keras/legacy/interfaces.py in wrapper(*args, **kwargs)
     89                 warnings.warn('Update your `' + object_name +
     90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)
---> 91             return func(*args, **kwargs)
     92         wrapper._original_function = func
     93         return wrapper

~/Code/tensorflow/lib/python3.6/site-packages/keras/engine/training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)
   2078 
   2079         do_validation = bool(validation_data)
-> 2080         self._make_train_function()
   2081         if do_validation:
   2082             self._make_test_function()

~/Code/tensorflow/lib/python3.6/site-packages/keras/engine/training.py in _make_train_function(self)
    988                     training_updates = self.optimizer.get_updates(
    989                         params=self._collected_trainable_weights,
--> 990                         loss=self.total_loss)
    991                 updates = self.updates + training_updates + self.metrics_updates
    992                 # Gets loss and metrics. Updates weights at each call.

~/Code/tensorflow/lib/python3.6/site-packages/keras/legacy/interfaces.py in wrapper(*args, **kwargs)
     89                 warnings.warn('Update your `' + object_name +
     90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)
---> 91             return func(*args, **kwargs)
     92         wrapper._original_function = func
     93         return wrapper

~/Code/tensorflow/lib/python3.6/site-packages/keras/optimizers.py in get_updates(self, loss, params)
    452 
    453         t = K.cast(self.iterations, K.floatx()) + 1
--> 454         lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
    455                      (1. - K.pow(self.beta_1, t)))
    456 

~/Code/tensorflow/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py in sqrt(x)
   1459     zero = _to_tensor(0., x.dtype.base_dtype)
   1460     inf = _to_tensor(np.inf, x.dtype.base_dtype)
-> 1461     x = tf.clip_by_value(x, zero, inf)
   1462     return tf.sqrt(x)
   1463 

~/Code/tensorflow/lib/python3.6/site-packages/tensorflow/python/ops/clip_ops.py in clip_by_value(t, clip_value_min, clip_value_max, name)
     58   """
     59   with ops.name_scope(name, "clip_by_value",
---> 60                       [t, clip_value_min, clip_value_max]) as name:
     61     t = ops.convert_to_tensor(t, name="t")
     62 

~/Code/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in __enter__(self)
   5614       if self._values is None:
   5615         self._values = []
-> 5616       g = _get_graph_from_inputs(self._values)
   5617       self._g_manager = g.as_default()
   5618       self._g_manager.__enter__()

~/Code/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _get_graph_from_inputs(op_input_list, graph)
   5282         graph = graph_element.graph
   5283       elif original_graph_element is not None:
-> 5284         _assert_same_graph(original_graph_element, graph_element)
   5285       elif graph_element.graph is not graph:
   5286         raise ValueError("%s is not from the passed-in graph." % graph_element)

~/Code/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _assert_same_graph(original_item, item)
   5218   if original_item.graph is not item.graph:
   5219     raise ValueError("%s must be from the same graph as %s." % (item,
-> 5220                                                                 original_item))
   5221 
   5222 

ValueError: Tensor("training/Adam/Const:0", shape=(), dtype=float32) must be from the same graph as Tensor("sub:0", shape=(), dtype=float32).

Steps to reproduce:

Just compile any model with an Adam optimizer and try to train it.

[Zhanchenhaha]

I also had the same error. How did you solve it? Thank you.

[pierluigiferrari]

I didn't. I just downgraded my TensorFlow version to 1.4.1, which is hardly a good solution.

[fchollet]

This is a user error. The stack trace means that the optimizer was first used in a TF graph A, and then reused in TF graph B. This can happen if you use the same optimizer after having called K.clear_session(), for instance.

What's your code?

[ymodak]

Closing this issue since its not a bug or feature request. Feel free to reopen if have any follow up questions. Thanks!

[P-mir]

I have the same problem, how to use the same optimizer in graph B without causing this error?