An error when saving/opening model and weights (babi_memnn)!

[dmikulic]

I just used the example script examples/babi_memnn.py and added the saving code at the end of the file.

json_string = answer.to_json()
open('my_model_architecture.json', 'w').write(json_string)
answer.save_weights('my_model_weights.h5', overwrite=True)

And after the training is finished I try to load the model with

from keras.models import model_from_json
model = model_from_json(open('my_model_architecture.json').read())
model.load_weights('my_model_weights.h5')

and then get the Exception: You are trying to load a weight file containing 11 layers into a model with 13 layers.

Here is the gist https://gist.github.com/anonymous/f433fa0b79d251b77edf8d862071841a

[fchollet]

Yes, there appears to be an issue with model serialization for deeply nested Sequential models. I'm looking into it.

Note that a simple fix would be to use the functional API (recommended for complex models) instead of deeply nested Sequential models.

[akmahaja]

@fchollet When I am using the functional API to create the model and try to save it using the following line, It gives me the following error saying:

  File "trainer.py", line 48, in create_training_features
  json_string = model.to_json()
  File "/usr/local/lib/python2.7/site-packages/keras/engine/topology.py", line 2368, in to_json
config = self.get_config()
 File "/usr/local/lib/python2.7/site-packages/keras/engine/topology.py", line 2163, in get_config
new_node_index = node_conversion_map[node_key]
KeyError: 'input_1_ib-0'

[fchollet]

Please provide the code to reproduce your error.

[akmahaja]

input1 = Input(shape=(maxlen,), dtype='int32',name='input1')
input2 = Input(shape=(maxlen,), dtype='int32',name='input2')
embedding = Embedding(input_dim=2000,
                         output_dim=300,
                         weights=weights)
e1=embedding(input1)
e2=embedding(input2)
dropout = Dropout(0.25)

f_rnn = LSTM(259, return_sequences=True, dropout_U=0.2, consume_less='mem')
b_rnn = LSTM(259, return_sequences=True, dropout_U=0.2, consume_less='mem',go_backwards=True)

input1_f_rnn = f_rnn(e1)
input1_b_rnn = b_rnn(e1)
input1_f_dropout = dropout(input1_f_rnn)
input1_b_dropout = dropout(input1_b_rnn)

# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False), output_shape=lambda x: (x[0], x[2]))
input1_pool = merge([maxpool(input1_f_dropout), maxpool(input1_b_dropout)], mode='concat', concat_axis=-1)

#input2
f_rnn = AttentionLSTM(259, input1_pool, single_attn=True, dropout_U=0.2,
                          return_sequences=True, consume_less='mem')
b_rnn = AttentionLSTM(259, input1_pool, single_attn=True, dropout_U=0.2,
                          return_sequences=True, consume_less='mem', go_backwards=True)
input2_f_rnn = f_rnn(e2)
input2_b_rnn = b_rnn(e2)
input2_pool = merge([maxpool(input2_f_rnn), maxpool(input2_b_rnn)], mode='concat', concat_axis=-1)

#couple of more dense layers
activation = Activation('tanh')
activated= activation(input2_pool)

dense_layer=Dense(15, activation='softmax',name='output')
output_final=dense_layer(activated)

model=Model(input=[input1,input2],output=[output_final])
model.compile(loss='categorical_crossentropy',optimizer='adam')

model.fit({'input1': X_train1, 'input2': X_train2},{'output': y_train},nb_epoch=1)

json_string=model.to_json()
open('model_architecture.json', 'w').write(json_string)
answer.save_weights('model_weights.h5', overwrite=True)

[joelthchao]

@akmahaja Is it your complete code? Your input2 is never used in the model, and remove input2 from Model(input=[input1, input2], ...) can solve the error.

[akmahaja]

@fchollet @joelthchao I am using the input2, I hadn't included that in the earlier version. I have updated the code with the latest version that I am working on, still the same issue.
AttentionLSTM that I am using in the above piece of code is taken from here:
https://github.com/codekansas/keras-language-modeling/blob/master/attention_lstm.py

[fchollet]

Please provide the code to reproduce your error.

The keyword here is "reproduce", if you share some code snippet that I can't run and thus doesn't allow me to reproduce your error, that's not very helpful...

[akmahaja]

@fchollet this is the entire code here that I am using to reproduce the error. Are you referring to X_train1 and X_train2, y_train values?
After experimentation issue is mainly because the "input1" is not being fed all the way to the output layer. It stops at "input1_pool" in the above code, hence when the overall network diagram is also created it does not show up there.

[fchollet]

I have fixed the issue from OP.

@akmahaja please open a thread for your own issue, which is different from this one. It should have a code snippet which I can copy and run (i.e. it should be self-contained, include imports and data. Use randomly generated data).