Weights by name in Keras

Question

After training a model using Keras, I can get a list of weight arrays using:

myModel.get_weights() 

or

myLayer.get_weights()

I'd like to know the names corresponding to each weight array. I know how to do this indirectly by saving the model and parsing the HDF5 file but surely there must be a direct way to accomplish this?

Answer 1

Function get_weights returns a list of numpy arrays with no name information in them.

As for Model.get_weights(), it's just the concatenation of Layer.get_weights() for each one of the [flattened] layers.

However, Layer.weights gives direct access to the backend variables, and these, yes, may have a name. The solution then is to iterate through each weight of each layer, retrieving its name attribute.

An example with VGG16:

from keras.applications.vgg16 import VGG16


model = VGG16()

names = [weight.name for layer in model.layers for weight in layer.weights]
weights = model.get_weights()

for name, weight in zip(names, weights):
    print(name, weight.shape)

which outputs:

block1_conv1_W_6:0 (3, 3, 3, 64)
block1_conv1_b_6:0 (64,)
block1_conv2_W_6:0 (3, 3, 64, 64)
block1_conv2_b_6:0 (64,)
block2_conv1_W_6:0 (3, 3, 64, 128)
block2_conv1_b_6:0 (128,)
block2_conv2_W_6:0 (3, 3, 128, 128)
block2_conv2_b_6:0 (128,)
block3_conv1_W_6:0 (3, 3, 128, 256)
block3_conv1_b_6:0 (256,)
block3_conv2_W_6:0 (3, 3, 256, 256)
block3_conv2_b_6:0 (256,)
block3_conv3_W_6:0 (3, 3, 256, 256)
block3_conv3_b_6:0 (256,)
block4_conv1_W_6:0 (3, 3, 256, 512)
block4_conv1_b_6:0 (512,)
block4_conv2_W_6:0 (3, 3, 512, 512)
block4_conv2_b_6:0 (512,)
block4_conv3_W_6:0 (3, 3, 512, 512)
block4_conv3_b_6:0 (512,)
block5_conv1_W_6:0 (3, 3, 512, 512)
block5_conv1_b_6:0 (512,)
block5_conv2_W_6:0 (3, 3, 512, 512)
block5_conv2_b_6:0 (512,)
block5_conv3_W_6:0 (3, 3, 512, 512)
block5_conv3_b_6:0 (512,)
fc1_W_6:0 (25088, 4096)
fc1_b_6:0 (4096,)
fc2_W_6:0 (4096, 4096)
fc2_b_6:0 (4096,)
predictions_W_6:0 (4096, 1000)
predictions_b_6:0 (1000,)