如何使用 dnnlib.tflib.network.Network 获得中间层 output

Question

I am trying StyleGAN2 NviLab，and I want to watch feature maps while generating, but how can I get the intermediate layer output? The model is loaded by dnnlib.tflib.network.Network and no document about this lib.

Answer 1

I had the exact same issue, with the exact same codebase. This is how I got intermediate outputs (TF 1.14.0):

Step 1:

Check the names of all layers, and their output shapes. This helps identify the name of the layer whose output you want:

network.print_layers()

Step 2:

Get a list of all layers in the network. Each element in this list is a tuple (layer name: string, layer output: tensor, layer variables)

layers = network.list_layers()

Step 3:

Get a reference to the tensor representing the layer's output, using the name of the layer whose output you need:

for layer in layers:
    print(layer[0], ' ', layer[1])
    if layer[0] == REQUIRED_LAYER_NAME:
        tensor = layer[1]

Finally, run the graph:

required_output = sess.run(tensor, feed_dict={'INPUT_TENSOR_NAME': input_numpy_array})