Are compiling and fitting required when using a CNN just to extract features?

I'm trying to implement a 1D CNN in Keras that will extract features from a spectrogram, as described in this paper: (p.170, section "1D Convolution for Audio Representation".

The goal is just for the spectrogram to "pass through" the CNN, which will "gradually aggregate" the 2D spectrogram to a 1D vector. (Later, this output is fused with features from a 2nd model and finally fed to an LSTM, but I'm just doing the CNN right now). There's no classification being performed; I'm just extracting the spectrogram features.

My question is: where do I stop the process in order to achieve this?

So far, I've built (a smaller version) of the CNN, as they described:

model = Sequential()    # Block 1  model.add(Conv1D(filters=32, kernel_size=5, input_shape=(timesteps, features), padding="same"))   model.add(BatchNormalization())  model.add(MaxPooling1D(pool_size=8))    # Block 2  model.add(Conv1D(filters=64, kernel_size=5))  model.add(BatchNormalization())  model.add(MaxPooling1D(pool_size=8))    # Get 1D vector  model.add(Flatten())    # What do I call next...?  

But from here, I don't know how to proceed. Specifically: do I call model.predict(spectrogram) directly, or do I need to .compile() and .fit() it first?

https://stackoverflow.com/questions/66573184/are-compiling-and-fitting-required-when-using-a-cnn-just-to-extract-features March 11, 2021 at 05:32AM