class Discriminator(Model):
def __init__(self, lr, hidden_units, reduce_units):
super(Discriminator, self).__init__()
self.opt = Adam(learning_rate=lr)
self.encoder = VAE_Encoder(latent_num=64)
self.concat = Concatenate()
self.reduces = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in reduce_units
]
# hidden
self.hiddens = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in hidden_units
]
# output
self.out = Dense(1,
activation='sigmoid',
kernel_initializer=tf.random_uniform_initializer(
minval=-3e-3, maxval=3e-3)) # 0: expert, 1: agent
def call(self, s, a):
s = self.encoder.forward(s)
for layer in self.reduces:
s = layer(s)
x = self.concat([s, a])
for layer in self.hiddens:
x = layer(x)
out = self.out(x)
return out
class Discriminator(Model):
def __init__(self, lr, hidden_units, reduce_units):
super(Discriminator, self).__init__()
self.opt = Adam(learning_rate=lr)
self.encoder = VAE_Encoder(latent_num=64)
self.concat = Concatenate()
self.reduces = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in reduce_units
]
# hidden
self.hiddens = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in hidden_units
]
self.latent_mu = Dense(LATENT_UNIT_NUM, activation='linear')
self.latent_std = Dense(LATENT_UNIT_NUM, activation='sigmoid')
# output
self.out = Dense(1,
activation='sigmoid',
kernel_initializer=tf.random_uniform_initializer(
minval=-3e-3, maxval=3e-3)) # 0: expert, 1: agent
def call(self, s, a):
mu, _, _ = self.encode(s, a)
out = self.decode(mu)
return out
def encode(self, s, a):
s = self.encoder.forward(s)
for layer in self.reduces:
s = layer(s)
x = self.concat([s, a])
for layer in self.hiddens:
x = layer(x)
mu = self.latent_mu(x)
std = self.latent_std(x)
# sampling
noise = std * np.random.normal(size=std.shape)
sampled = noise + mu
return mu, std, sampled
def decode(self, latent):
return self.out(latent)
class Critic(Model):
def __init__(self, lr, hidden_units):
super(Critic, self).__init__()
self.opt = Adam(learning_rate=lr)
self.encoder = VAE_Encoder(latent_num=64)
# hidden
self.hiddens = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in hidden_units
]
# output
self.value = Dense(1,
kernel_initializer=tf.random_uniform_initializer(
minval=-3e-3, maxval=3e-3))
def call(self, s):
x = self.encoder.forward(s)
for layer in self.hiddens:
x = layer(x)
value = self.value(x)
return value
class Actor(Model):
def __init__(self, lr, hidden_units):
super(Actor, self).__init__()
self.opt = Adam(learning_rate=lr)
self.encoder = VAE_Encoder(latent_num=64)
# hidden
self.hiddens = [
Dense(unit, activation='relu', kernel_initializer='he_normal')
for unit in hidden_units
]
# output
self.out = Dense(ACTION_NUM,
activation='softmax',
kernel_initializer=tf.random_uniform_initializer(
minval=-3e-3, maxval=3e-3))
def call(self, s):
x = self.encoder.forward(s)
for layer in self.hiddens:
x = layer(x)
policy = self.out(x)
return policy