def __init__(
self,
learning_rate,
num_layers,
size,
size_layer,
output_size,
kernel_size=3,
n_attn_heads=16,
dropout=0.9,
):
self.X = tf.placeholder(tf.float32, (None, None, size))
self.Y = tf.placeholder(tf.float32, (None, output_size))
encoder_embedded = tf.layers.dense(self.X, size_layer)
encoder_embedded += position_encoding(encoder_embedded)
e = tf.identity(encoder_embedded)
for i in range(num_layers):
dilation_rate = 2**i
pad_sz = (kernel_size - 1) * dilation_rate
with tf.variable_scope('block_%d' % i):
encoder_embedded += cnn_block(encoder_embedded, dilation_rate,
pad_sz, size_layer, kernel_size)
encoder_output, output_memory = encoder_embedded, encoder_embedded + e
g = tf.identity(encoder_embedded)
for i in range(num_layers):
dilation_rate = 2**i
pad_sz = (kernel_size - 1) * dilation_rate
with tf.variable_scope('decode_%d' % i):
attn_res = h = cnn_block(encoder_embedded, dilation_rate,
pad_sz, size_layer, kernel_size)
C = []
for j in range(n_attn_heads):
h_ = tf.layers.dense(h, size_layer // n_attn_heads)
g_ = tf.layers.dense(g, size_layer // n_attn_heads)
zu_ = tf.layers.dense(encoder_output,
size_layer // n_attn_heads)
ze_ = tf.layers.dense(output_memory,
size_layer // n_attn_heads)
d = tf.layers.dense(h_, size_layer // n_attn_heads) + g_
dz = tf.matmul(d, tf.transpose(zu_, [0, 2, 1]))
a = tf.nn.softmax(dz)
c_ = tf.matmul(a, ze_)
C.append(c_)
c = tf.concat(C, 2)
h = tf.layers.dense(attn_res + c, size_layer)
h = tf.nn.dropout(h, keep_prob=dropout)
encoder_embedded += h
encoder_embedded = tf.sigmoid(encoder_embedded[-1])
self.logits = tf.layers.dense(encoder_embedded, output_size)
self.cost = tf.reduce_mean(tf.square(self.Y - self.logits))
self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(
self.cost)
def __init__(self, state_size, window_size, trend, skip):
self.state_size = state_size
self.window_size = window_size
self.half_window = window_size // 2
self.trend = trend
self.skip = skip
tf.reset_default_graph()
self.X = tf.placeholder(tf.float32, (None, self.state_size))
self.Y = tf.placeholder(tf.float32, (None, self.state_size))
self.ACTION = tf.placeholder(tf.float32, (None))
self.REWARD = tf.placeholder(tf.float32, (None))
self.batch_size = tf.shape(self.ACTION)[0]
with tf.variable_scope('curiosity_model'):
action = tf.reshape(self.ACTION, (-1,1))
state_action = tf.concat([self.X, action], axis=1)
save_state = tf.identity(self.Y)
feed = tf.layers.dense(state_action, 32, activation=tf.nn.relu)
self.curiosity_logits = tf.layers.dense(feed, self.state_size)
self.curiosity_cost = tf.reduce_sum(tf.square(save_state - self.curiosity_logits), axis=1)
self.curiosity_optimizer = tf.train.RMSPropOptimizer(self.LEARNING_RATE) .minimize(tf.reduce_mean(self.curiosity_cost))
total_reward = tf.add(self.curiosity_cost, self.REWARD)
with tf.variable_scope("q_model"):
with tf.variable_scope("eval_net"):
x_action = tf.layers.dense(self.X, 128, tf.nn.relu)
self.logits = tf.layers.dense(x_action, self.OUTPUT_SIZE)
with tf.variable_scope("target_net"):
y_action = tf.layers.dense(self.Y, 128, tf.nn.relu)
y_q = tf.layers.dense(y_action, self.OUTPUT_SIZE)
q_target = total_reward + self.GAMMA * tf.reduce_max(y_q, axis=1)
action = tf.cast(self.ACTION, tf.int32)
action_indices = tf.stack([tf.range(self.batch_size, dtype=tf.int32), action], axis=1)
q = tf.gather_nd(params=self.logits, indices=action_indices)
self.cost = tf.losses.mean_squared_error(labels=q_target, predictions=q)
self.optimizer = tf.train.RMSPropOptimizer(self.LEARNING_RATE).minimize(
self.cost, var_list=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "q_model/eval_net"))
t_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='q_model/target_net')
e_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='q_model/eval_net')
self.target_replace_op = [tf.assign(t, e) for t, e in zip(t_params, e_params)]
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
def __init__(
self,
learning_rate,
num_layers,
size,
size_layer,
output_size,
kernel_size=3,
n_attn_heads=16,
dropout=0.9,
):
self.X = tf.placeholder(tf.float32, (None, None, size))
self.Y = tf.placeholder(tf.float32, (None, output_size))
encoder_embedded = tf.layers.dense(self.X, size_layer)
e = tf.identity(encoder_embedded)
for i in range(num_layers):
z = layer(
encoder_embedded,
encoder_block,
kernel_size,
size_layer * 2,
encoder_embedded,
)
z = tf.nn.dropout(z, keep_prob=dropout)
encoder_embedded = z
encoder_output, output_memory = z, z + e
g = tf.identity(encoder_embedded)
for i in range(num_layers):
attn_res = h = layer(
encoder_embedded,
decoder_block,
kernel_size,
size_layer * 2,
residual=tf.zeros_like(encoder_embedded),
)
C = []
for j in range(n_attn_heads):
h_ = tf.layers.dense(h, size_layer // n_attn_heads)
g_ = tf.layers.dense(g, size_layer // n_attn_heads)
zu_ = tf.layers.dense(encoder_output,
size_layer // n_attn_heads)
ze_ = tf.layers.dense(output_memory,
size_layer // n_attn_heads)
d = tf.layers.dense(h_, size_layer // n_attn_heads) + g_
dz = tf.matmul(d, tf.transpose(zu_, [0, 2, 1]))
a = tf.nn.softmax(dz)
c_ = tf.matmul(a, ze_)
C.append(c_)
c = tf.concat(C, 2)
h = tf.layers.dense(attn_res + c, size_layer)
h = tf.nn.dropout(h, keep_prob=dropout)
encoder_embedded = h
encoder_embedded = tf.sigmoid(encoder_embedded[-1])
self.logits = tf.layers.dense(encoder_embedded, output_size)
self.cost = tf.reduce_mean(tf.square(self.Y - self.logits))
self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(
self.cost)
def __init__(self, state_size, window_size, trend, skip):
self.state_size = state_size
self.window_size = window_size
self.half_window = window_size // 2
self.trend = trend
self.skip = skip
tf.reset_default_graph()
self.INITIAL_FEATURES = np.zeros((4, self.state_size))
self.X = tf.placeholder(tf.float32, (None, None, self.state_size))
self.Y = tf.placeholder(tf.float32, (None, None, self.state_size))
self.hidden_layer = tf.placeholder(tf.float32,
(None, 2 * self.LAYER_SIZE))
self.ACTION = tf.placeholder(tf.float32, (None))
self.REWARD = tf.placeholder(tf.float32, (None))
self.batch_size = tf.shape(self.ACTION)[0]
self.seq_len = tf.shape(self.X)[1]
with tf.variable_scope('curiosity_model'):
action = tf.reshape(self.ACTION, (-1, 1, 1))
repeat_action = tf.tile(action, [1, self.seq_len, 1])
state_action = tf.concat([self.X, repeat_action], axis=-1)
save_state = tf.identity(self.Y)
cell = tf.nn.rnn_cell.LSTMCell(self.LAYER_SIZE,
state_is_tuple=False)
self.rnn, last_state = tf.nn.dynamic_rnn(
inputs=state_action,
cell=cell,
dtype=tf.float32,
initial_state=self.hidden_layer)
self.curiosity_logits = tf.layers.dense(self.rnn[:, -1],
self.state_size)
self.curiosity_cost = tf.reduce_sum(
tf.square(save_state[:, -1] - self.curiosity_logits), axis=1)
self.curiosity_optimizer = tf.train.RMSPropOptimizer(
self.LEARNING_RATE).minimize(
tf.reduce_mean(self.curiosity_cost))
total_reward = tf.add(self.curiosity_cost, self.REWARD)
with tf.variable_scope("q_model"):
with tf.variable_scope("eval_net"):
cell = tf.nn.rnn_cell.LSTMCell(self.LAYER_SIZE,
state_is_tuple=False)
rnn, self.last_state = tf.nn.dynamic_rnn(
inputs=self.X,
cell=cell,
dtype=tf.float32,
initial_state=self.hidden_layer)
self.logits = tf.layers.dense(rnn[:, -1], self.OUTPUT_SIZE)
with tf.variable_scope("target_net"):
cell = tf.nn.rnn_cell.LSTMCell(self.LAYER_SIZE,
state_is_tuple=False)
rnn, last_state = tf.nn.dynamic_rnn(
inputs=self.Y,
cell=cell,
dtype=tf.float32,
initial_state=self.hidden_layer)
y_q = tf.layers.dense(rnn[:, -1], self.OUTPUT_SIZE)
q_target = total_reward + self.GAMMA * tf.reduce_max(y_q, axis=1)
action = tf.cast(self.ACTION, tf.int32)
action_indices = tf.stack(
[tf.range(self.batch_size, dtype=tf.int32), action], axis=1)
q = tf.gather_nd(params=self.logits, indices=action_indices)
self.cost = tf.losses.mean_squared_error(labels=q_target,
predictions=q)
self.optimizer = tf.train.RMSPropOptimizer(
self.LEARNING_RATE).minimize(
self.cost,
var_list=tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, "q_model/eval_net"))
t_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='q_model/target_net')
e_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='q_model/eval_net')
self.target_replace_op = [
tf.assign(t, e) for t, e in zip(t_params, e_params)
]
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())