Compile GAN

loan_data_compile_gan.py

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input,Dense,BatchNormalization,Concatenate,GaussianNoise
from tensorflow.keras.optimizers import Nadam

#define latent dimension size
latent_dim = int(np.ceil(np.log(len(train_df)*len(data))))

#function for building generator network
def compile_generator():
    
    #input for random seed vecotors
    latent_inputs = Input(shape=(latent_dim,),name='latent')
    #input for target specification
    target_inputs = Input(shape=(2,),name='target')
    
    inputs = {'latent':latent_inputs,'target':target_inputs}
    
    net = Concatenate()([latent_inputs,target_inputs])
    
    for _ in range(2):
        net = Dense(32+len(data), activation='relu',
                    kernel_initializer='he_uniform')(net)
        net = BatchNormalization()(net)
    
    outputs = {}
    
    #numeric data outputs
    for name in numeric_data:
        outputs[name] = Dense(1,activation='tanh',
                              kernel_initializer='glorot_uniform',name=name)(net)
    
    string_nets = Dense(len(string_data),activation='relu',
                        kernel_initializer='he_uniform')(net)
    string_nets = BatchNormalization()(string_nets)
    
    #string data outputs
    for name,n_token in n_tokens.items():
        string_net = Dense(n_embeddings[name],activation='relu',
                           kernel_initializer='he_uniform')(string_nets)
        string_net = BatchNormalization()(string_net)
        outputs[name] = Dense(n_token,activation='softmax',
                              kernel_initializer='glorot_uniform',name=name)(net)
    
    generator = Model(inputs=inputs, outputs=outputs)
    generator.compile(loss='categorical_crossentropy',
                      optimizer=Nadam(clipnorm=1.))

    return generator


#function for building discriminator network
def compile_discriminator():
    
    inputs = {}
    
    numeric_nets = []
    string_nets = []
    
    #numerica data inputs
    for name in numeric_data:
        numeric_input = Input(shape=(1,),name=name)
        inputs[name] = numeric_input
        
        numeric_net = GaussianNoise(0.01)(numeric_input)
        numeric_nets.append(numeric_net)
    
    #string data inputs
    for name,n_token in n_tokens.items():
        string_input = Input(shape=(n_token,),name=name)
        inputs[name] = string_input
    
        string_net = GaussianNoise(0.05)(string_input)
        string_net = Dense(n_embeddings[name],activation='relu',
                           kernel_initializer='he_uniform')(string_net)
        string_nets.append(string_net)
        
    string_nets = Concatenate()(string_nets)
    string_nets = BatchNormalization()(string_nets)
    string_nets = [Dense(len(string_data),activation='relu',
                         kernel_initializer='he_uniform')(string_nets)]
    
    net = Concatenate()(numeric_nets + string_nets)
    net = BatchNormalization()(net)

    for _ in range(2):
        net = Dense(64+len(data), activation='relu',
                    kernel_initializer='he_uniform')(net)
        net = BatchNormalization()(net)
    
    #discrimination/classification
    outputs = Dense(3, activation='softmax',
                    kernel_initializer='glorot_uniform')(net)
    
    discriminator = Model(inputs=inputs, outputs=outputs)
    discriminator.compile(loss='categorical_crossentropy',
                          optimizer=Nadam(clipnorm=1.))

    return discriminator


#function for building GAN network
def compile_gan():                       
    
    #disable discriminator training
    discriminator.trainable = False
    
    #input for random seed vecotors
    latent_inputs = Input(shape=(latent_dim,),name='latent')
    #input for target specification
    target_inputs = Input(shape=(2,),name='target')
    
    inputs = {'latent':latent_inputs,'target':target_inputs}

    net = generator([latent_inputs, target_inputs])
    outputs = discriminator(net)

    gan = Model(inputs=inputs, outputs=outputs)
    gan.compile(loss='categorical_crossentropy', optimizer=Nadam(clipnorm=1.))

    return gan
    

#compile GAN network
generator = compile_generator()
discriminator = compile_discriminator()
gan = compile_gan()