def create_model(self, model_info):
"""Create Deep-Q network."""
state = Input(shape=self.state_dim)
denselayer = Dense(HIDDEN_SIZE, activation='relu')(state)
for _ in range(NUM_LAYERS - 1):
denselayer = Dense(HIDDEN_SIZE, activation='relu')(denselayer)
value = Dense(self.action_dim, activation='linear')(denselayer)
if self.dueling:
adv = Dense(1, activation='linear')(denselayer)
mean = Lambda(layer_normalize)(value)
value = Lambda(layer_add)([adv, mean])
model = Model(inputs=state, outputs=value)
adam = Adam(lr=self.learning_rate)
model.compile(loss='mse', optimizer=adam)
self.infer_state = tf.placeholder(tf.float32,
name="infer_input",
shape=(None, ) +
tuple(self.state_dim))
self.infer_v = model(self.infer_state)
self.actor_var = TFVariables([self.infer_v], self.sess)
self.sess.run(tf.initialize_all_variables())
return model
def create_model(self, model_info):
"""Create keras model."""
state_input = Input(shape=self.state_dim, name='state_input')
advantage = Input(shape=(1, ), name='adv')
denselayer = Dense(HIDDEN_SIZE, activation='relu')(state_input)
for _ in range(NUM_LAYERS - 1):
denselayer = Dense(HIDDEN_SIZE, activation='relu')(denselayer)
out_actions = Dense(self.action_dim,
activation='softmax',
name='output_actions')(denselayer) # y_pred
out_value = Dense(1, name='output_value')(denselayer)
model = Model(inputs=[state_input, advantage],
outputs=[out_actions, out_value])
losses = {
"output_actions": impala_loss(advantage),
"output_value": 'mse'
}
lossweights = {"output_actions": 1.0, "output_value": .5}
model.compile(optimizer=Adam(lr=LR),
loss=losses,
loss_weights=lossweights)
self.infer_state = tf.placeholder(tf.float32,
name="infer_state",
shape=(None, ) +
tuple(self.state_dim))
self.adv = tf.placeholder(tf.float32, name="adv", shape=(None, 1))
self.infer_p, self.infer_v = model([self.infer_state, self.adv])
self.actor_var = TFVariables([self.infer_p, self.infer_v], self.sess)
self.sess.run(tf.initialize_all_variables())
return model
def build_graph(self, input_type, model):
# pylint: disable=W0201
self.state_ph = tf.placeholder(input_type,
name='state',
shape=(None, *self.state_dim))
self.old_logp_ph = tf.placeholder(tf.float32,
name='old_log_p',
shape=(None, 1))
self.adv_ph = tf.placeholder(tf.float32,
name='advantage',
shape=(None, 1))
self.old_v_ph = tf.placeholder(tf.float32,
name='old_v',
shape=(None, 1))
self.target_v_ph = tf.placeholder(tf.float32,
name='target_value',
shape=(None, 1))
pi_latent, self.out_v = model(self.state_ph)
if self.action_type == 'Categorical':
self.behavior_action_ph = tf.placeholder(tf.int32,
name='behavior_action',
shape=(None, ))
dist_param = pi_latent
elif self.action_type == 'DiagGaussian':
# fixme: add input dependant log_std logic
self.behavior_action_ph = tf.placeholder(tf.float32,
name='real_action',
shape=(None,
self.action_dim))
log_std = tf.get_variable('pi_logstd',
shape=(1, self.action_dim),
initializer=tf.zeros_initializer())
dist_param = tf.concat([pi_latent, pi_latent * 0.0 + log_std],
axis=-1)
else:
raise NotImplementedError(
'action type: {} not match any implemented distributions.'.
format(self.action_type))
self.dist.init_by_param(dist_param)
self.action = self.dist.sample()
self.action_log_prob = self.dist.log_prob(self.action)
self.actor_var = TFVariables([self.action_log_prob, self.out_v],
self.sess)
self.actor_loss = actor_loss_with_entropy(self.dist, self.adv_ph,
self.old_logp_ph,
self.behavior_action_ph,
self.clip_ratio,
self.ent_coef)
self.critic_loss = critic_loss(self.target_v_ph, self.out_v,
self.old_v_ph, self.vf_clip)
self.loss = self.actor_loss + self.critic_loss_coef * self.critic_loss
self.train_op = self.build_train_op(self.loss)
self.sess.run(tf.initialize_all_variables())
def create_model(self, model_info):
"""Create Deep-Q CNN network."""
state = Input(shape=self.state_dim, dtype="int8")
state1 = Lambda(lambda x: K.cast(x, dtype='float32') / 255.)(state)
convlayer = Conv2D(32, (8, 8),
strides=(4, 4),
activation='relu',
padding='valid')(state1)
convlayer = Conv2D(64, (4, 4),
strides=(2, 2),
activation='relu',
padding='valid')(convlayer)
convlayer = Conv2D(64, (3, 3),
strides=(1, 1),
activation='relu',
padding='valid')(convlayer)
flattenlayer = Flatten()(convlayer)
denselayer = Dense(256, activation='relu')(flattenlayer)
value = Dense(self.action_dim, activation='linear')(denselayer)
if self.dueling:
adv = Dense(1, activation='linear')(denselayer)
mean = Lambda(layer_normalize)(value)
value = Lambda(layer_add)([adv, mean])
model = Model(inputs=state, outputs=value)
adam = Adam(lr=self.learning_rate, clipnorm=10.)
model.compile(loss='mse', optimizer=adam)
if model_info.get("summary"):
model.summary()
self.infer_state = tf.placeholder(tf.int8,
name="infer_input",
shape=(None, ) +
tuple(self.state_dim))
self.infer_v = model(self.infer_state)
self.actor_var = TFVariables([self.infer_v], self.sess)
self.sess.run(tf.initialize_all_variables())
return model
def create_model(self, model_info):
"""Create Deep-Q network."""
user_input = Input(shape=(self.user_dim,), name="user_input", dtype=self.input_type)
history_click_input = Input(
shape=(self.n_history_click * self.item_dim), name="history_click",
dtype=self.input_type
)
history_no_click_input = Input(
shape=(self.n_history_no_click * self.item_dim), name="history_no_click",
dtype=self.input_type
)
item_input = Input(shape=(self.item_dim,), name="item_input", dtype=self.input_type)
shared_embedding = Embedding(
self.vocab_size,
self.emb_dim,
name="Emb",
mask_zero=True,
embeddings_initializer=self.embedding_initializer,
trainable=False,
) # un-trainable
gru_click = GRU(self.item_dim * self.emb_dim)
gru_no_click = GRU(self.item_dim * self.emb_dim)
user_feature = Flatten()(shared_embedding(user_input))
item_feature = Flatten()(shared_embedding(item_input))
history_click_feature = Reshape(
(self.n_history_click, self.item_dim * self.emb_dim)
)(shared_embedding(history_click_input))
history_click_feature = gru_click(history_click_feature)
history_no_click_feature = Reshape(
(self.n_history_no_click, self.item_dim * self.emb_dim)
)(shared_embedding(history_no_click_input))
history_no_click_feature = gru_no_click(history_no_click_feature)
x = concatenate(
[
user_feature,
history_click_feature,
history_no_click_feature,
item_feature,
]
)
x_dense1 = Dense(128, activation="relu")(x)
x_dense2 = Dense(128, activation="relu")(x_dense1)
# ctr_pred = Dense(1, activation="linear", name="q_value")(x_dense2)
ctr_pred = Dense(1, activation=self.last_act, name="q_value")(x_dense2)
model = Model(
inputs=[
user_input,
history_click_input,
history_no_click_input,
item_input,
],
outputs=ctr_pred,
)
model.compile(loss="mse", optimizer=Adam(lr=self.learning_rate))
if self._summary:
model.summary()
self.user_input = tf.placeholder(
dtype=self.input_type, name="user_input", shape=(None, self.user_dim)
)
self.history_click_input = tf.placeholder(
dtype=self.input_type,
name="history_click_input",
shape=(None, self.n_history_click * self.item_dim),
)
self.history_no_click_input = tf.placeholder(
dtype=self.input_type,
name="history_no_click_input",
shape=(None, self.n_history_no_click * self.item_dim),
)
self.item_input = tf.placeholder(
dtype=self.input_type, name="item_input", shape=(None, self.item_dim)
)
self.ctr_predict = model(
[
self.user_input,
self.history_click_input,
self.history_no_click_input,
self.item_input,
]
)
self.actor_var = TFVariables([self.ctr_predict], self.sess)
self.sess.run(tf.initialize_all_variables())
return model
def build_graph(self):
self.build_train_graph()
self.build_infer_graph()
self.sess.run(tf.initialize_all_variables())