Errors in saving a Keras model structure

[Ming-Hua Chung]

Hi, all,

While following the instruction in FAQ to save a model structure, I kept getting error massages regardless whether I save it in JSON or YAML. The error messages are the same on both. Does anyone have the same problem? If so, how do I get around it? Appreciate your help!

# save as JSON

vae_json = vae.to_json() # save as YAML vae_yaml = vae.to_yaml()

vae2 = model_from_json(vae_json)

vae2 = model_from_yaml(vae_yaml)

Here is the error messages:

>>> vae2 = model_from_yaml(vae_yaml)

Traceback (most recent call last):
  File "<input>", line 1, in <module>
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/models.py", line 27, in model_from_yaml
    return layer_from_config(config, custom_objects=custom_objects)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/utils/layer_utils.py", line 35, in layer_from_config
    return layer_class.from_config(config['config'])
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 2216, in from_config
    layer(input_tensors)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 485, in __call__
    self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 543, in add_inbound_node
    Node.create_node(self, inbound_layers, node_indices, tensor_indices)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 153, in create_node
    output_tensors = to_list(outbound_layer.call(input_tensors, mask=input_masks))
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/layers/core.py", line 446, in call
    return self.function(x, **arguments)
  File "<input>", line 22, in sampling
NameError: global name 'batch_size' is not defined

Michael

[Ming-Hua Chung]

Hi, Constantin,

I have follows the autoencoder example in Keras blog to set up my model:

batch_size = 16

original_dim = 784

latent_dim = 2

intermediate_dim = 128

epsilon_std = 0.01

nb_epoch = 40

x = Input(batch_shape=(batch_size, original_dim))

h = Dense(intermediate_dim, activation='relu')(x)

And this is how my model eventually looks like:

>>> vae.summary()

____________________________________________________________________________________________________

Layer (type)                       Output Shape        Param #     Connected to

=======================================================================================

input_1 (InputLayer)               (16, 784)           0

____________________________________________________________________________________________________

dense_1 (Dense)                    (16, 128)           100480      input_1[0][0]

____________________________________________________________________________________________________

dense_2 (Dense)                    (16, 2)             258         dense_1[0][0]

____________________________________________________________________________________________________

dense_3 (Dense)                    (16, 2)             258         dense_1[0][0]

____________________________________________________________________________________________________

lambda_1 (Lambda)                  ((16, 2), 2)        0           dense_2[0][0]

                                                                                  dense_3[0][0] 

And this is how I train it:

vae.fit(x_train, x_train,

        shuffle=True,

        nb_epoch=nb_epoch,

        batch_size=batch_size,

        validation_data=(x_test, x_test))

I don't think I miss anything since the model coding part is basically a direct copy from Keras blog example. But thank you for your reply anyway.

Michael

On Thursday, June 2, 2016 at 3:01:53 AM UTC-5, Constantin Ei wrote:

Hi,

seems to me, that you simply have to define "batch_size". Your error message says, it is required but missing.

-- Constantin

[Ming-Hua Chung]

After further digging, I realize that the problem is probably caused by function from_config since I got the same error message when I try to load back my model configuration. Still no solution though.

>>> config = vae.get_config()

>>> vae = Model.from_config(config)

Traceback (most recent call last):
  File "<input>", line 1, in <module>

  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 2216, in from_config
    layer(input_tensors)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 485, in __call__
    self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 543, in add_inbound_node
    Node.create_node(self, inbound_layers, node_indices, tensor_indices)
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/engine/topology.py", line 153, in create_node
    output_tensors = to_list(outbound_layer.call(input_tensors, mask=input_masks))
  File "/usr/local/miniconda/lib/python2.7/site-packages/keras/layers/core.py", line 446, in call
    return self.function(x, **arguments)
  File "<input>", line 22, in sampling
NameError: global name 'batch_size' is not defined

>>> batch_size

16

[Ming-Hua Chung]

Ok, so I solve the problem myself. In case anyone is interested, here is the solution.

Here is the codes of variational autoencoder from Keras Blog:

def sampling(args):
    z_mean, z_log_sigma = args
    epsilon = K.random_normal(shape=(batch_size, latent_dim),
                              mean=0., std=epsilon_std)
    return z_mean + K.exp(z_log_sigma) * epsilon

Once I changed to this (below), then I don't get the error messages anymore.

def sampling(args):
    z_mean, z_log_sigma = args
    epsilon = K.random_normal(shape=(16, 2),
                              mean=0., std=0.01)
    return z_mean + K.exp(z_log_sigma) * epsilon

So, it is not really from_config, but from the sampling function.

Michael

[zgb...@gmail.com]

That helps, but it still cannot solve the question. 

If we save the variational autoencoder model: 

autoencoder = vae.save('variationalAutoencoder.h5')

and then load it in another script: 

vae = load_model('variationalAutoencoder.h5')

The errors will occur, like

batch_size not difined, latent_dim not defined, invalid objects vae_loss

Anybody can solve it?