Keras: RuntimeError: Unable to create link (Name already exists) : Bidirectionnal name ?

[emicol]

I'm up-to-date with the master branch of Keras, and have the same kind of problem than here : https://github.com/fchollet/keras/issues/new

When I save my model, equiped with a bidirectional layer, that appears :
autoencoder.autoencoder.save("data/autoencoder.h5")
File "/usr/local/lib/python3.5/dist-packages/keras/engine/topology.py", line 2425, in save
save_model(self, filepath, overwrite)
File "/usr/local/lib/python3.5/dist-packages/keras/models.py", line 109, in save_model
topology.save_weights_to_hdf5_group(model_weights_group, model_layers)
File "/usr/local/lib/python3.5/dist-packages/keras/engine/topology.py", line 2717, in save_weights_to_hdf5_group
dtype=val.dtype)
File "/usr/local/lib/python3.5/dist-packages/h5py/_hl/group.py", line 108, in create_dataset
self[name] = dset
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/tmp/pip-at6d2npe-build/h5py/_objects.c:2684)
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/tmp/pip-at6d2npe-build/h5py/_objects.c:2642)
File "/usr/local/lib/python3.5/dist-packages/h5py/_hl/group.py", line 277, in __setitem__
h5o.link(obj.id, self.id, name, lcpl=lcpl, lapl=self._lapl)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/tmp/pip-at6d2npe-build/h5py/_objects.c:2684)
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/tmp/pip-at6d2npe-build/h5py/_objects.c:2642)
File "h5py/h5o.pyx", line 202, in h5py.h5o.link (/tmp/pip-at6d2npe-build/h5py/h5o.c:3731)
RuntimeError: Unable to create link (Name already exists)

It was functional before the update.

A code like that won't produce a stored model:

    from keras.layers import Input, Dense, LSTM, Embedding, Bidirectional
    from keras.models import Model

    batch_size = 256
    max_features_1 = 256 
    max_sequence_1 = 58 
    max_sequence_2 = 40 

    max_len = 58
    # this is the size of our encoded representations    
    encoding_dim = 40  #

    input_word = Input(shape=(max_sequence_1,))
    embed = Embedding(max_features_1, output_dim=48, input_length=max_sequence_1)(input_word)
    be1 = Bidirectional(LSTM(20, return_sequences=True))(embed)  
    be2 = Bidirectional(LSTM(20))(be1)  # 20 le nb de neurones -
    encoded = Dense(encoding_dim, activation='relu')(be2)


    # "decoded" is the "lossy" reconstruction of the input
    decoded = Dense(len(get_dictionnaire("", False)), activation='sigmoid')(encoded)

    # this model maps an input to its reconstruction
    autoencoder = Model(inputs=input_word, outputs=decoded)

    # this model maps an input to its encoded representation
    encoder = Model(inputs=input_word, outputs=encoded)

    # create a placeholder for an encoded (40-dimensional) input
    encoded_input = Input(shape=(encoding_dim,))
    # retrieve the last layer of the autoencoder model
    decoder_layer = autoencoder.layers[-1]
    # create the decoder model
    decoder = Model(inputs=encoded_input, outputs=decoder_layer(encoded_input))

    autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
    autoencoder.save("test.h5")

[around1991]

I also get the same issue.

[vadimnazarov]

Same issue for me too.

[srph25]

Confirmed, fails with ModelCheckpoint also.

[martin-favre]

Also have the same issue with Bidirectional. Same issue seem to arise with Timedistributed.

[drewlevin]

Confirmed Bidirectional LSTM using ModelCheckpoint.

[matt-gardner]

FYI, #5939 has a fix for this. At least it works for me.

[ylmeng]

I still have the same problem with a custom recurrent layer. It works if I do not use Bidirectional.
I suspect it is a name issue but have not figured it out.