Graph disconnected error when using skip connections in an autoencoder

Question

I have implemented a simple variational autoencoder in Keras with 2 convolutional layers in the encoder and decoder. The code is shown below. Now, I have extended my implementation with two skip connections (similar to U-Net). The skip connections are named merge1 and merge2 in the below code. Without the skip connections everything works fine but with the skip connections I'm getting the following error message:

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("encoder_input:0", shape=(?, 64, 80, 1), dtype=float32) at layer "encoder_input". The following previous layers were accessed without issue: []

Is there a problem in my code?

import keras
from keras import backend as K
from keras.layers import (Dense, Input, Flatten)
from keras.layers import Conv2D, Lambda, MaxPooling2D, UpSampling2D, concatenate
from keras.models import Model
from keras.layers import Reshape
from keras.losses import mse

def sampling(args):
    z_mean, z_log_var = args
    batch = K.shape(z_mean)[0]
    dim = K.int_shape(z_mean)[1]
    epsilon = K.random_normal(shape=(batch, dim))
    return z_mean + K.exp(0.5 * z_log_var) * epsilon

image_size = (64,80,1)
inputs = Input(shape=image_size, name='encoder_input')

conv1 = Conv2D(64, 3, activation='relu', padding='same')(inputs)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(128, 3, activation='relu', padding='same')(pool1)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)

shape = K.int_shape(pool2)

x = Flatten()(pool2)
x = Dense(16, activation='relu')(x)
z_mean = Dense(6, name='z_mean')(x)
z_log_var = Dense(6, name='z_log_var')(x)

z = Lambda(sampling, output_shape=(6,), name='z')([z_mean, z_log_var])
encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')

latent_inputs = Input(shape=(6,), name='z_sampling')
x = Dense(16, activation='relu')(latent_inputs)
x = Dense(shape[1] * shape[2] * shape[3], activation='relu')(x)
x = Reshape((shape[1], shape[2], shape[3]))(x)

up1 = UpSampling2D((2, 2))(x)
up1 = Conv2D(128, 2, activation='relu', padding='same')(up1)
merge1 = concatenate([conv2, up1], axis=3)

up2 = UpSampling2D((2, 2))(merge1)
up2 = Conv2D(64, 2, activation='relu', padding='same')(up2)
merge2 = concatenate([conv1, up2], axis=3)

out = Conv2D(1, 1, activation='sigmoid')(merge2)

decoder = Model(latent_inputs, out, name='decoder')

outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs, name='vae')

def vae_loss(x, x_decoded_mean):
    reconstruction_loss = mse(K.flatten(x), K.flatten(x_decoded_mean))
    reconstruction_loss *= image_size[0] * image_size[1]
    kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
    kl_loss = K.sum(kl_loss, axis=-1)
    kl_loss *= -0.5
    vae_loss = K.mean(reconstruction_loss + kl_loss)
    return vae_loss

optimizer = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.000)
vae.compile(loss=vae_loss, optimizer=optimizer)
vae.fit(train_X, train_X,
        epochs=500,
        batch_size=128,
        verbose=1,
        shuffle=True,
        validation_data=(valid_X, valid_X))

Answer 1

Your decoder takes conv1 and conv2 as inputs. It cannot be created with simply Model(latent_inputs, ...)

You need Model([inputs, latent_inputs], ...)