deeplearning.py 文件源码

def gen_Model(num_units, actfn='linear', reg_coeff=0.0, last_act='softmax'):
    ''' Generate a neural network model of approporiate architecture
    Args:
        num_units: architecture of network in the format [n1, n2, ... , nL]
        actfn: activation function for hidden layers ('relu'/'sigmoid'/'linear'/'softmax')
        reg_coeff: L2-regularization coefficient
        last_act: activation function for final layer ('relu'/'sigmoid'/'linear'/'softmax')
    Output:
        model: Keras sequential model with appropriate fully-connected architecture
    '''
    model = Sequential()
    for i in range(1, len(num_units)):
        if i == 1 and i < len(num_units) - 1:
            model.add(Dense(input_dim=num_units[0], output_dim=num_units[i], activation=actfn,
                W_regularizer=Reg.l2(l=reg_coeff), init='glorot_normal'))
        elif i == 1 and i == len(num_units) - 1:
            model.add(Dense(input_dim=num_units[0], output_dim=num_units[i], activation=last_act,
                W_regularizer=Reg.l2(l=reg_coeff), init='glorot_normal'))
        elif i < len(num_units) - 1:
            model.add(Dense(output_dim=num_units[i], activation=actfn,
                W_regularizer=Reg.l2(l=reg_coeff), init='glorot_normal'))
        elif i == len(num_units) - 1:
            model.add(Dense(output_dim=num_units[i], activation=last_act,
                W_regularizer=Reg.l2(l=reg_coeff), init='glorot_normal'))
    return model