Keras: Have saved weights of entire model: How to extract out individual layer weights?

[dc-ai]

With some help, I was able to setup a NN within a NN.

Basically the architecture is as follows:

• There is a main neural network that uses pretrained weights that are frozen
• There is a sub neural network that you are training
• However because the sub neural network is incorporated in the main network as a "model", when you do model.summary(), you do not see the layers of the sub network. Consequently, when the main network is saved, the weights of the sub neural network is not saved as individual layers.
• When I need to do transfer learning and use part of the sub neural network, I can't do it, because the weights are saved to the entire sub network, and not each layer in the sub network.

Has anyone encountered these type of situations before?

In code:

Step 1: Setup sub NN within main NN and train

# Main NN: Simple convolutional autoencoder
input_img = Input(shape=(1, 80, 80), name="input_layer")
intermediate = Convolution2D(32, 3, 3, name="conv1")(input_img)
decoded = Convolution2D(10, 3, 3, name="conv2")(intermediate)

fc_flatten = Flatten()(intermediate)

# Sub NN: Uses output from encoding part to do supervised learning on FC layers
input_NN2 = Input( (None,80*80*32))
y = Dense(64, name="dense3")(input_NN2)
fc_output = Dense(1)(y)
model2 = Model([input_NN2], [fc_output], name="submodel")

fc_output = model2([fc_flatten])
model = Model(input=[input_img], output=[decoded, fc_output])
model.compile(loss=["mse","mse"],optimizer="adam")

Step 2: Setup modified NN that has parts of the main NN and sub NN and load weights

The major difficulty is that the weights of the layer "dense3" is not stored as a layer in the previous step. Instead all the weights are saved in the model, "submodel". I can't figure out how to "separate out" the weights of all layers inside "submodel".

input_img = Input(shape=(1, 80, 80), name="input_layer")
intermediate = Convolution2D(32, 3, 3, name="conv1")(input_img)
fc_flatten = Flatten()(intermediate)
y = Dense(64, name="dense3")(fc_flatten)
new_output = Dense(4)(y)

newmodel = Model(input=[input_img], output=[new_output])
newmodel.compile(loss=["mse"],optimizer="adam")

I tried opening the .hdf5 file, but that format is not readable in text. Any suggestions? Thanks!

[unrealwill]

Using save_weights(), and load_weights() works.
You can also use model.get_layer( either by name or index) to get the layer and you can then call get_weights() or set_weights(). And do the mapping by matching names.

Alternatively when the structure of the network is known :
Usually when you build a network, you can return a tuple of model
(combined_network, subnetwork1, subnetwork2 ) which all share the same shared variables

If you need to transfer the weights you to another network or subnetwork for which you have a tuple.
(new_combined_network, new_subnetwork1, new_subnetwork2)

doing new_combined_network.set_weights( combined_network.get_weights() ) will set all the weights including the subnetworks.

doing new_subnetwork1.set_weights( subnetwork1.get_weights()) will only transfer the weight of the subnetwork1