Understanding Sharing Keras Layers

[nikcheerla]

I'm trying to build a variational autoencoder that encodes and decodes music sequence data (of shape (50, 84)). Simultaneously, I want to predict the next element inside the music sequence (vector of shape (84,)). So I need to reuse a trainable decoder that maps from the latent space to the output space – for both encoding and next element prediction.

autoencode_output = TimeDistributed(decoder) (z)

x = GRU(64, dropout_W=0.1, dropout_U=0.1, return_sequences=True) (z)
x = GRU(12, dropout_W=0.1, dropout_U=0.1, return_sequences=False) (x)
next_output = decoder(x)

model = Model(input_img, [autoencode_output, next_output])

This is the relevant part of the code; I reuse the "decoder" model twice (in the TimeDistributed wrapper and also by itself). I keep on getting a 'this shared variable already has an update expression' error, however. When do these errors happen when reusing models, and what's the cleanest way to fix them?

Thanks!

[unrealwill]

Hello, I tried to reproduce your bug but couldn't : def buildDecoderModel(): inp = Input(batch_shape=(None,100)) decout = Dense(10)(inp) decoder = Model([inp],[decout] ) inp2 = Input(batch_shape=(None,30,100)) emb = TimeDistributed( decoder )(inp2) gru = GRU(100,return_sequences=True)(emb) gru2 = GRU(100,return_sequences=False)(gru) dec = decoder(gru2) model = Model( [inp2],[gru,dec]) model.compile( loss=["mse","mse"], optimizer="adam" ) return model

I tested with theano :

import numpy as np m= buildDecoderModel() m.fit(np.random.randn(10,30,100),[np.zeros((10,30,100)) , np.zeros((10,10)) ] )

It works fine.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.