def create_model(self, model_info):
state_input = Input(shape=self.state_dim, name='state_input', dtype='uint8')
state_input_1 = Lambda(layer_function)(state_input)
advantage = Input(shape=(1, ), name='adv')
convlayer = Conv2D(32, (8, 8), strides=(4, 4), activation='relu', padding='valid')(state_input_1)
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)
out_actions = Dense(self.action_dim, activation='softmax', name='output_actions')(denselayer)
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}
decay_value = 0.00000000512
model.compile(optimizer=Adam(lr=LR, clipnorm=40., decay=decay_value), loss=losses, loss_weights=lossweights)
self.infer_state = tf.placeholder(tf.uint8, 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.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 create_dyn_network(self):
conditioned_hidden = Input(shape=HIDDEN_OUT + self.action_dim)
hidden = Dense(256, activation='relu')(conditioned_hidden)
hidden = Dense(128, activation='relu')(hidden)
out_h = Dense(HIDDEN_OUT, activation='relu')(hidden)
out_r = Dense(self.reward_support_size, activation='softmax')(hidden)
return Model(inputs=conditioned_hidden, outputs=[out_h, out_r])
def get_mlp_backbone(state_dim,
act_dim,
hidden_sizes,
activation,
vf_share_layers=False,
summary=False):
"""Get mlp backbone."""
state_input = Input(shape=state_dim, name='obs')
if not vf_share_layers:
dense_layer_pi = bulid_mlp_layers(state_input, hidden_sizes,
activation, 'pi')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer_pi)
dense_layer_v = bulid_mlp_layers(state_input, hidden_sizes, activation,
'v')
out_value = Dense(1, activation=None,
name='output_value')(dense_layer_v)
else:
dense_layer = bulid_mlp_layers(state_input, hidden_sizes, activation,
'shared')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer)
out_value = Dense(1, activation=None, name='output_value')(dense_layer)
model = Model(inputs=[state_input], outputs=[pi_latent, out_value])
if summary:
model.summary()
return model
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_dyn_network(self):
conditioned_hidden = Input(shape=HIDDEN2_UNITS + self.action_dim)
hidden = Dense(HIDDEN1_UNITS, activation='relu')(conditioned_hidden)
out_h = Dense(HIDDEN2_UNITS, activation='relu')(hidden)
# out_h = Lambda(hidden_normlize)(out_h)
out_r = Dense(self.reward_support_size, activation='softmax')(hidden)
return Model(inputs=conditioned_hidden, outputs=[out_h, out_r])
def create_model(self, model_info):
"""Create Deep-Q CNN network."""
state = Input(shape=self.state_dim, dtype="uint8")
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)
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.uint8,
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):
state_input = Input(shape=self.state_dim, name='state_input')
advantage = Input(shape=(1, ), name='adv')
old_prediction = Input(shape=(self.action_dim, ), name='old_p')
old_value = Input(shape=(1, ), name='old_v')
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)
out_value = Dense(1, name='output_value')(denselayer)
model = Model(inputs=[state_input], outputs=[out_actions, out_value])
if model_info.get("summary"):
model.summary()
self.build_graph(tf.float32, model)
return model
def create_rep_network(self):
obs = Input(shape=self.state_dim, name='rep_input')
obs_1 = Lambda(lambda x: tf.cast(x, dtype='float32') / 255.)(obs)
convlayer = Conv2D(32, (8, 8), strides=(4, 4), activation='relu', padding='valid')(obs_1)
convlayer = Conv2D(32, (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(HIDDEN_OUT, activation='relu')(flattenlayer)
# hidden = Lambda(hidden_normlize)(denselayer)
hidden = denselayer
return Model(inputs=obs, outputs=hidden)
def create_model(self, model_info):
"""method for creating DQN 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)
model = Model(inputs=state, outputs=value)
adam = Adam(lr=self.learning_rate)
model.compile(loss='mse', optimizer=adam)
return model
def get_cnn_backbone(state_dim,
act_dim,
hidden_sizes,
activation,
filter_arches,
vf_share_layers=True,
summary=False,
dtype='uint8'):
"""Get CNN backbone."""
state_input_raw = Input(shape=state_dim, name='obs')
if dtype == 'uint8':
state_input = Lambda(layer_function)(state_input_raw)
elif dtype == 'float32':
state_input = state_input_raw
else:
raise ValueError(
'dtype: {} not supported automatically, please implement it yourself'
.format(dtype))
if vf_share_layers:
conv_layer = build_conv_layers(state_input, filter_arches, activation,
'shared')
flatten_layer = Flatten()(conv_layer)
dense_layer = bulid_mlp_layers(flatten_layer, hidden_sizes, activation,
'shared')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer)
out_value = Dense(1, activation=None, name='output_value')(dense_layer)
else:
conv_layer_pi = build_conv_layers(state_input, filter_arches,
activation, 'pi')
conv_layer_v = build_conv_layers(state_input, filter_arches,
activation, 'v')
flatten_layer_pi = Flatten()(conv_layer_pi)
flatten_layer_v = Flatten()(conv_layer_v)
dense_layer_pi = bulid_mlp_layers(flatten_layer_pi, hidden_sizes,
activation, 'pi')
dense_layer_v = bulid_mlp_layers(flatten_layer_v, hidden_sizes,
activation, 'v')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer_pi)
out_value = Dense(1, activation=None,
name='output_value')(dense_layer_v)
model = Model(inputs=[state_input_raw], outputs=[pi_latent, out_value])
if summary:
model.summary()
return model
def create_model(self, model_info):
state_input = Input(shape=self.state_dim, name='state_input', dtype='uint8')
state_input_1 = Lambda(layer_function)(state_input)
convlayer = Conv2D(32, (8, 8), strides=(4, 4), activation='relu', padding='valid')(state_input_1)
convlayer = Conv2D(32, (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', name='dense_1')(flattenlayer)
out_actions = Dense(self.action_dim, activation='softmax', name='output_actions_raw')(denselayer)
out_value = Dense(1, name='output_value')(denselayer)
model = Model(inputs=[state_input], outputs=[out_actions, out_value])
self.build_graph(np.uint8, model)
return model
def get_cnn_backbone(state_dim,
act_dim,
hidden_sizes,
activation,
filter_arches,
vf_share_layers=True,
summary=False):
"""Get CNN backbone."""
state_input_raw = Input(shape=state_dim, name='obs')
state_input = Lambda(layer_function)(state_input_raw)
if vf_share_layers:
conv_layer = build_conv_layers(state_input, filter_arches, activation,
'shared')
flatten_layer = Flatten()(conv_layer)
dense_layer = bulid_mlp_layers(flatten_layer, hidden_sizes, activation,
'shared')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer)
out_value = Dense(1, activation=None, name='output_value')(dense_layer)
else:
conv_layer_pi = build_conv_layers(state_input, filter_arches,
activation, 'pi')
conv_layer_v = build_conv_layers(state_input, filter_arches,
activation, 'v')
flatten_layer_pi = Flatten()(conv_layer_pi)
flatten_layer_v = Flatten()(conv_layer_v)
dense_layer_pi = bulid_mlp_layers(flatten_layer_pi, hidden_sizes,
activation, 'pi')
dense_layer_v = bulid_mlp_layers(flatten_layer_v, hidden_sizes,
activation, 'v')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer_pi)
out_value = Dense(1, activation=None,
name='output_value')(dense_layer_v)
model = Model(inputs=[state_input_raw], outputs=[pi_latent, out_value])
if summary:
model.summary()
return model
def get_mlp_backbone(state_dim,
act_dim,
hidden_sizes,
activation,
vf_share_layers=False,
summary=False,
dtype='float32'):
"""Get mlp backbone."""
state_input_raw = Input(shape=state_dim, name='obs')
if dtype == 'float32':
state_input = state_input_raw
else:
raise ValueError(
'dtype: {} not supported automatically, please implement it yourself'
.format(dtype))
if not vf_share_layers:
dense_layer_pi = bulid_mlp_layers(state_input, hidden_sizes,
activation, 'pi')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer_pi)
dense_layer_v = bulid_mlp_layers(state_input, hidden_sizes, activation,
'v')
out_value = Dense(1, activation=None,
name='output_value')(dense_layer_v)
else:
dense_layer = bulid_mlp_layers(state_input, hidden_sizes, activation,
'shared')
pi_latent = Dense(act_dim, activation=None,
name='pi_latent')(dense_layer)
out_value = Dense(1, activation=None, name='output_value')(dense_layer)
model = Model(inputs=[state_input], outputs=[pi_latent, out_value])
if summary:
model.summary()
return model
def create_policy_network(self):
hidden_input = Input(shape=HIDDEN2_UNITS, name='hidden_input')
hidden = Dense(HIDDEN1_UNITS, activation='relu')(hidden_input)
out_v = Dense(self.value_support_size, activation='softmax')(hidden)
out_p = Dense(self.action_dim, activation='softmax')(hidden)
return Model(inputs=hidden_input, outputs=[out_p, out_v])
def create_rep_network(self):
obs = Input(shape=self.state_dim, name='rep_input')
hidden = Dense(HIDDEN1_UNITS, activation='relu')(obs)
out_rep = Dense(HIDDEN2_UNITS, activation='relu')(hidden)
# out_rep = Lambda(hidden_normlize)(out_rep)
return Model(inputs=obs, outputs=out_rep)
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