How to resize input size for Conv1DTranspose layer in Keras?

Question

I'm using Keras to build GAN based on Conv1DTranspose layers. I would like to implement a GAN model (especially for the generator). However, I couldn't make the right architecture. Could you help me for these points:

(1) The problem is the size of my input (noise) does not match as per Keras documentation. So, I couldn't build the network from the beginning.

(2) How to perform "reshape" from Conv1DTranspose layer output to Conv1D layer ?

Thank you,,

The generator architecture which I want to replicate:

This is my code:

# define generator model
def define_generator(latent_dim, n_outputs=1):
    model = Sequential()

    # reshape layer
    model.add(Dense(1 * 10 * 10, activation="relu", input_dim=(latent_dim)))
    model.add(Reshape((10, 10, 1)))

    # 1D Transposed Convolutional Layer
    model.add(Conv1DTranspose(filters=32, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))

    model.add(Conv1DTranspose(filters=64, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))

    model.add(Conv1DTranspose(filters=128, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))

    # 1D Convolutional Layer
    model.add(Conv1D(filters=128, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))

    model.add(Conv1D(filters=64, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))

    model.add(Conv1D(filters=32, kernel_size=4, strides=1))
    model.add(BatchNormalization(momentum=0.8))
    model.add(LeakyReLU(alpha=0.1))
    
    # output layer
    model.add(Dense(n_outputs, activation='sigmoid'))

    print('Generator')
    model.summary()

    return model

The error messages:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-70-a8709ad0f7a6> in <module>()
      1 latent_dim = 100
      2 
----> 3 generator = define_generator(latent_dim)

6 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/layers/convolutional.py in build(self, input_shape)
    944     if len(input_shape) != 3:
    945       raise ValueError('Inputs should have rank 3. Received input shape: ' +
--> 946                        str(input_shape))
    947     channel_axis = self._get_channel_axis()
    948     if input_shape.dims[channel_axis].value is None:

ValueError: Inputs should have rank 3. Received input shape: (None, 10, 10, 1)

Answer 1

the Conv1DTranspose return a tensor with shape (None, 10, 10, 1), so before the 1D Convolutional layer try to add one Reshape layer to squeeze it back to 3D tensor with the appropriate shape