TF 2 - Keras - Merging two seperately trained models to a new ensemble model - artificial-intelligence

Question

Newbie here ;)
I need your help ;)
I have following problem:
I want to merge two TF 2.0 - Keras Models into a new Model.
Expect for example a ResNet50V2 without the head, and retrained on new data - saved weights are provided and loaded:
resnet50v2 = ResNet50V2(weights='imagenet', include_top=False)
#model.summary()
last_layer = resnet50v2.output
x = GlobalAveragePooling2D()(last_layer)
x = Dense(768, activation='relu', name='img_dense_768')(x)
out = Dense(NUM_CLASS, activation='softmax', name='img_output_layer')(x)
resnet50v2 = Model(inputs=resnet50v2.input, outputs=out)
resnet50v2.load_weights(pathToImageModel)
Expect for example a Bert Network retrained on new data - saved weights are provided, in the same way as shown before.
Now I want to skip the last softmax layer from the two models and "merge" them into a new Model.
So I took the layer before the last layer, which names I know:
model_image_output = model_image.get_layer('img_dense_768').output
model_bert_output = model_bert.get_layer('bert_output_layer_768').output
I built two new models with these informations:
model_img = Model(model_image.inputs, model_image_output)
model_bert = Model(model_bert.inputs, model_bert_output)
So now I want to concatenate the outputs of the two models into a new one and add a new "head"
concatenate = Concatenate()([model_img.output,model_bert.output])
x = Dense(512, activation = 'relu')(concatenate)
x = Dropout(0.4)(x)
output = Dense(NUM_CLASS, activation='softmax', name='final_multimodal_output')(x)
model_concat = Model([model_image.input, model_bert.input])
So far so good, model code seemed to be valid but then my knowledge ends.
The main questions are:
Also if the last softmax layers are skipped, the loaded weights should be still available or ?
The new concatened model wants to be build, ok but this ends in this error message, which I only partially understand:
NotImplementedError: When subclassing the Model class, you should implement a call method.
Is there another way to create for example the whole ensemble network and load only the pretrained parts of it and leave the rest beeing trainable?
Am I missing something? I never did subclassing the model class? If so, it was not intended XD
May I ask kindly for some hints ?
Thanks in advance!!
Kind regards ragitagha
UPDATE:
So to find my mistake more quickly and provide the solution in a more precise way:
I had to change the lines
concatenate = Concatenate()([model_img.output,model_bert.output])
x = Dense(512, activation = 'relu')(concatenate)
x = Dropout(0.4)(x)
output = Dense(NUM_CLASS, activation='softmax', name='final_multimodal_output')(x)
model_concat = Model([model_image.input, model_bert.input])
to:
concatenate = Concatenate()([model_img.output,model_bert.output])
x = Dense(512, activation = 'relu')(concatenate)
x = Dropout(0.4)(x)
output = Dense(NUM_CLASS, activation='softmax', name='final_multimodal_output')(x)
model_concat = Model([model_image.input, model_bert.input], output)