Is there any way to change a connection in a pre-trained tensorflow model and use them while keeping the all other layers unchanged?

General Discussion
, ,
#1

Is there any way to change a connection in a pre-trained tensorflow model and use them while keeping the all other layers above, in-between and below them unchanged?
Say I have a pre-trained model and I want to change a specific connection.
Say a model has:
`

 Input
   |
   v
 layer1
   |
   v
 layer2 ----------------------
   |                         |
   v                         |
 layer3 -----------          |
   |              |          |
   v              |          |
 layer4           |          |
   |              |          |
   v              |          |
 layer5 <----------          |
   |                         |
   v                         |
 layer6 <---------------------
   |
   v
 Dense

`
They are all connected but there is also a skip connection between:

  • layer2 and layer6
  • layer3 and layer5

Now I want to insert a new layer say layer4_1 on the skip connection between layer3 and layer5
So it becomes:

 Input
   |
   v
 layer1
   |
   v
 layer2 ----------------------
   |                         |
   v                         |
 layer3 -----------          |
   |              |          |
   v              v          |
 layer4       layer4_1       |
   |              |          |
   v              |          |
 layer5 <----------          |
   |                         |
   v                         |
 layer6 <---------------------
   |
   v
 Dense

The rest of the connections remain the same even the connection between layer2 and layer6.

I tried the following but it doesn’t work:

def mymodel(input_shape, n_classes):
    pretrained_model = net(input_shape, n_classes)
    desired_layer = pretrained_model.get_layer('block_13_expand_relu')   # layer3
    desired_layer_output = desired_layer.output
    
    new_layer = NewLayer(top=2)         # layer4_1
    new_output = new_layer(desired_layer_output)
    merged_output = new_output
    
    upsampling_layer = pretrained_model.get_layer("decoder_stage0_upsampling")  # layer4
    upsampling_layer_output = upsampling_layer.output
    
    merged_output = tf.keras.layers.Concatenate()([upsampling_layer_output, merged_output])   # tried changing layer5 as couldn't do it otherwise
    
    decoder_conv_layer = pretrained_model.get_layer("decoder_stage0_concat")      # layer6
    new_model = Model(inputs=decoder_conv_layer.output, outputs=pretrained_model.output)
    
    output = new_model(merged_output)
    
    model = Model(inputs=pretrained_model.input, outputs=output)
    
    return model

Sorry can’t give original pretrained model structure as it is too big.

Can someone please help me?

The error message I am receiving is:

ValueError: Found input tensor cannot be reached given provided
output tensors. Please make sure the tensor
KerasTensor(type_spec=TensorSpec(shape=(None, 256, 256, 3),
dtype=tf.float32, name='input_1'), name='input_1', description="created
by layer 'input_1'") is included in the model inputs when building
functional model.
#3

Try referencing each layer individually and reconstruct the model with your modifications ?
If the model is too big to do so, have you tried tf.keras.layers.Add() in the skip connection ?