dcgans_amc.py 文件源码

def generator_model_bpsk(no_bits_in_a_frame):
    """
    BPSK outputs will be generated by CCN.
    CCN would be 1-2x because x is binary and the output should be bipolar. 
    Also, it is 1-tap processing. For 16-QAM, it will be more compliated. 
    I should consider how to optimize stride or oversampling/max polling 
    in a network. For GANs, hyperparameters can be more well optimized than 
    conventional feedforward networks. 
    While I was watching RNN-LSTM, I realized that many hyperparameters such as
    gating variables are optimized by networks itself. Those values have been optimized 
    by grid search or some other external techniques. However, RNN can do it by itself online. 
    These capability may come from RNN superpower. Similarly, many hyperparameters can be
    easily optimized in GANs. 
    """
    model = Sequential()
    model.add(Convolution1D(
        1, 1,
        input_shape=(no_bits_in_a_frame, 1)))
    return model