Here is an example to solve similar questions from the issue #43561
When I was trying to load the sequential model here using tf.keras.models.load_model in TF 2.3.1, an error is thrown at the following location:
~/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/functional.py in _should_skip_first_node(layer)
1031 return (isinstance(layer, Functional) and
1032 # Filter out Sequential models without an input shape.
-> 1033 isinstance(layer._layers[0], input_layer_module.InputLayer))
1034
1035
IndexError: list index out of range
The model is believed to be trained using keras and under TF1.9, and the model definition can be found here, and here's the code for training.
Then I downgraded to TF 2.2 and 2.1 with the same code above, it threw the error just as #35934 Keras Model Errors on Loading - 'list' object has no attribute 'items'
Then I downgraded to TF 2.0, the code was executing indefinitely. Finally I had to manually stop it:
/opt/conda/lib/python3.6/site-packages/tensorflow_core/python/pywrap_tensorflow_internal.py in IsMapping(o)
2569
2570 """
-> 2571 return _pywrap_tensorflow_internal.IsMapping(o)
2572
2573 def IsMappingView(o):
KeyboardInterrupt:
Then I have tried to use keras instead of tf.keras with TF 2.3.1 and Keras 2.3.1, first I encountered an error that can be solved in this way: https://github.com/tensorflow/tensorflow/issues/38589#issuecomment-665930503 . Then another error occurs:
~/.local/lib/python3.7/site-packages/tensorflow/python/keras/backend.py in function(inputs, outputs, updates, name, **kwargs)
3931 if updates:
3932 raise ValueError('`updates` argument is not supported during '
-> 3933 'eager execution. You passed: %s' % (updates,))
3934 from tensorflow.python.keras import models # pylint: disable=g-import-not-at-top
3935 from tensorflow.python.keras.utils import tf_utils # pylint: disable=g-import-not-at-top
ValueError: `updates` argument is not supported during eager execution. You passed: [<tf.Variable 'UnreadVariable' shape=() dtype=int64, numpy=0>, <tf.Variable 'UnreadVariable' shape=(3, 3, 3, 32) dtype=float32, numpy=
array([[[[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0.],
......
So this way fails.
Solutions
One way is to use TF 1.15.4 and Keras 2.3.1, and finally it worked out fine, inputs, outputs, summary etc. are all parsed correctly, as well as being able to run data through the model.
Another is to modify the TF 2.3.1 source code so that the model can be used in latest version using tensorflow keras. You have to redefining _should_skip_first_node in file tensorflow/python/keras/engine/functional.py:
def _should_skip_first_node(layer):
"""Returns True if the first layer node should not be saved or loaded."""
# Networks that are constructed with an Input layer/shape start with a
# pre-existing node linking their input to output. This node is excluded from
# the network config.
if layer._layers:
return (isinstance(layer, Functional) and
# Filter out Sequential models without an input shape.
isinstance(layer._layers[0], input_layer_module.InputLayer))
else:
return isinstance(layer, Functional)
Afterwards
I have submitted a PR #43570 to tensorflow, and it get fixed in Tensorflow 2.5.0.
Top comments (0)