Keras: Save Model: TypeError: ('Not JSON Serializable:', Dimension(2048))

[Offpics]

I made a stackoverflow question about this because it seems people had similar problem in the past https://stackoverflow.com/questions/48691449/typeerror-not-json-serializable-dimension2048.

When I try to save this particular model it gives TypeError: ('Not JSON Serializable:', Dimension(2048)).

I can save other models without a problem and I don't understand why this one doesn't work.

I tried to save it on Windows 10 with python_ver = 3.6, tensorflow_ver = 1.6-rc0 and Ubuntu 16.04 with python_ver = 3.6, tensorflow_ver = 1.3.

I created a model and trained it using code below.

from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import InputLayer
from tensorflow.python.keras.layers import Dense


# Declare variables for model.
transfer_len = 2048
num_classes = 3


# Model creation.
model = Sequential()
# Input layer of shape 2048.
model.add(InputLayer(input_shape = (transfer_len,)))
# Fully connected 1024.
model.add(Dense(1024, activation='relu'))
# Output layer.
model.add(Dense(num_classes, activation='softmax'))


from tensorflow.python.keras.optimizers import Adam

optimizer = Adam(lr=1e-3)

model.compile(optimizer = optimizer,
             loss = 'categorical_crossentropy',
             metrics=['accuracy'])

model.fit(x = transfer_values_train,
         y = labels_train,
         epochs = 20, batch_size = 100, verbose=0)

When I try to save it it gives following error.

output_path = "model.keras"
model.save(output_path)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-22-6a252d3d7102> in <module>()
----> 1 model.save(output_path)

~\Anaconda3\envs\gpu\lib\site-packages\tensorflow\python\keras\_impl\keras\engine\topology.py in save(self, filepath, overwrite, include_optimizer)
   1044     """
   1045     from tensorflow.python.keras._impl.keras.models import save_model  # pylint: disable=g-import-not-at-top
-> 1046     save_model(self, filepath, overwrite, include_optimizer)
   1047 
   1048   def save_weights(self, filepath, overwrite=True):

~\Anaconda3\envs\gpu\lib\site-packages\tensorflow\python\keras\_impl\keras\models.py in save_model(model, filepath, overwrite, include_optimizer)
    131             'config': model.get_config()
    132         },
--> 133         default=get_json_type).encode('utf8')
    134 
    135     model_weights_group = f.create_group('model_weights')

~\Anaconda3\envs\gpu\lib\json\__init__.py in dumps(obj, skipkeys, ensure_ascii, check_circular, allow_nan, cls, indent, separators, default, sort_keys, **kw)
    236         check_circular=check_circular, allow_nan=allow_nan, indent=indent,
    237         separators=separators, default=default, sort_keys=sort_keys,
--> 238         **kw).encode(obj)
    239 
    240 

~\Anaconda3\envs\gpu\lib\json\encoder.py in encode(self, o)
    197         # exceptions aren't as detailed.  The list call should be roughly
    198         # equivalent to the PySequence_Fast that ''.join() would do.
--> 199         chunks = self.iterencode(o, _one_shot=True)
    200         if not isinstance(chunks, (list, tuple)):
    201             chunks = list(chunks)

~\Anaconda3\envs\gpu\lib\json\encoder.py in iterencode(self, o, _one_shot)
    255                 self.key_separator, self.item_separator, self.sort_keys,
    256                 self.skipkeys, _one_shot)
--> 257         return _iterencode(o, 0)
    258 
    259 def _make_iterencode(markers, _default, _encoder, _indent, _floatstr,

~\Anaconda3\envs\gpu\lib\site-packages\tensorflow\python\keras\_impl\keras\models.py in get_json_type(obj)
    113       return obj.__name__
    114 
--> 115     raise TypeError('Not JSON Serializable:', obj)
    116 
    117   from tensorflow.python.keras._impl.keras import __version__ as keras_version  # pylint: disable=g-import-not-at-top

TypeError: ('Not JSON Serializable:', Dimension(2048)

[Offpics]

Tried with tensorflow_ver = 1.5.0, python_ver = 3.5.0 and it gives the same error.
Also changed the model to Functional and raises the same error.

inputs = Input(shape=(transfer_len,))

net = inputs

net = Dense(1024, activation='relu')(net)

net = Dense(num_classes, activation='softmax')(net)

outputs = net

[Offpics]

Okay, so the transfer_len variable was a type 'tensorflow.python.framework.tensor_shape.Dimension'.
Changed to int and it saves normally.
Now the question is wheter it is possible to fix it in the save method.

[ghost]

Hello!

I was able to fix this by changing get_json_type here:
tensorflow/python/keras/_impl/keras/engine/saving.py.

This is the function that's called from the \tensorflow\python\keras\_impl\keras\models.py file that raised your error.

Looks like the original function essentially if/else's the classes to find a JSON-friendly representation and raises a TypeError if it can't.

The following lines makes it happy, but I'm sure there's a much better, more principled fix. Anyone know what the best place to put code like this would be? I am more than happy to take a stab at it, just wondering where it would be at.
It seems like the standalone Keras fixed this when it marshal.dump(...)s the object, not sure what the tf.Keras equivalent is.

# ... earlier get_json_type code
# NOTE: Hacky fix to serialize Dimension objects.
from tensorflow.python.framework.tensor_shape import Dimension
if type(obj) == Dimension:
  return int(obj)
# original error raised here
raise TypeError('Not JSON Serializable:', obj)

[ArchieMeng]

Since Dimension object can be None like 'Dimension(None)'.A better fix can be:

from tensorflow.python.framework.tensor_shape import Dimension
if type(obj) == Dimension:
  return int(obj.value or 0)

[ialhashim]

I still need to add the hacks suggested by @clck @ArchieMeng . Why isn't this issue fixed?

[gillesdegottex]

@clck @ArchieMeng definitely works. it seems get_json_type() is in engine/network.py though.

The question is how to make from tensorflow.python.framework.tensor_shape import Dimension backend-independent. Which might explain why there has been no PR about this.

[cottrell]

Possibly related: https://github.com/keras-team/keras/issues/12473

[sandippaul7204]

I am having the same problem. I have added the code lines in Keras as suggested by you. It returns the same error.

[ArchieMeng]

I am having the same problem. I have added the code lines in Keras as suggested by you. It returns the same error.

Well. I just went back to keras with Deep learning recently, and I used the model.save method as well. However, it doesn't raise the same error as I got many years ago, and it works pretty well.
So, how did you create your model?