Dilations and nb_stack relationship

[rhaghi]

Hi, First, thanks a lot for making this package. It is very beneficial to me. I am trying to make a NN model with the following code, but I face some issues I don't understand. I want the dilation to be one single number. Something like dilations = (2,). The code is:

    input_dim = 9
    i = Input(batch_shape=(None, time_steps, input_dim))
    d1 = (1,)
    d2 = (2,)
    d3 = (4,)

    nbf =2

    ks1 = 128
    ks2 = 64
    ks3 = 32
    nbs =1 
    tcn1 = TCN(nb_filters=nbf,
           kernel_size=ks1,
           dilations=d1,
           return_sequences=True, 
           nb_stacks=nbs,
           padding='same')(I)
    tcn2 = TCN(nb_filters=nbf,
           kernel_size=ks2,
           dilations=d2,
           return_sequences=True,
           nb_stacks=nbs, 
           padding='same')(tcn1)
    tcn3 = TCN(nb_filters=nbf,
           kernel_size=ks3,
           dilations=d3,
           return_sequences=True,
           nb_stacks=nbs, 
           padding='same')(tcn2)
    o = Dense(1,activation='linear')(tcn3)
    m = Model(inputs=[i], outputs=[o])
    m.compile(optimizer='adam', loss='mse')

    history = m.fit(x,y, batch_size=12, epochs=300, validation_split=0.2, shuffle=True, callbacks=None)"

And then the error is:

ValueError: Exception encountered when calling layer "tcn_56" (type TCN).

in user code:

File "/usr/local/lib/python3.7/dist-packages/tcn/tcn.py", line 341, in call  *
    x = layers.add(self.skip_connections, name='Add_Skip_Connections')
File "/usr/local/lib/python3.7/dist-packages/keras/layers/merge.py", line 791, in add  **
    return Add(**kwargs)(inputs)
File "/usr/local/lib/python3.7/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.7/dist-packages/keras/layers/merge.py", line 92, in build
    raise ValueError('A merge layer should be called '

ValueError: A merge layer should be called on a list of at least 2 inputs. Got 1 inputs. Full input_shape received: ListWrapper([(None, 1200, 2)])

Call arguments received:
  • inputs=tf.Tensor(shape=(None, 1200, 9), dtype=float32)
  • training=False
  • kwargs=<class 'inspect._empty'>

If I change nb_stacks = 2, the code will work without any problem (the loss is high, but it runs!) What am I overlooking?

[philipperemy]

@rhaghi I think it's a bug. TCN does not seem to work with a dilation of just one value. There's a merge layer somewhere that is problematic. I'll see how to fix that.

[philipperemy]

For the loss, I guess it's normal that it's high at the beginning if you stack layers.

[philipperemy]

That one should fix it: https://github.com/philipperemy/keras-tcn/pull/238.

[philipperemy]

Pushed in 3.5.0.