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 _build(self, inputs, prev_state):
"""Connects the DNC core into the graph.
Args:
inputs: Tensor input.
prev_state: A `DNCState` tuple containing the fields `access_output`,
`access_state` and `controller_state`. `access_state` is a 3-D Tensor
of shape `[batch_size, num_reads, word_size]` containing read words.
`access_state` is a tuple of the access module's state, and
`controller_state` is a tuple of controller module's state.
Returns:
A tuple `(output, next_state)` where `output` is a tensor and `next_state`
is a `DNCState` tuple containing the fields `access_output`,
`access_state`, and `controller_state`.
"""
prev_access_output = prev_state.access_output
prev_access_state = prev_state.access_state
prev_controller_state = prev_state.controller_state
batch_flatten = snt.BatchFlatten()
controller_input = tf.concat(
[batch_flatten(inputs),
batch_flatten(prev_access_output)], 1)
controller_output, controller_state = self._controller(
controller_input, prev_controller_state)
controller_output = self._clip_if_enabled(controller_output)
controller_state = snt.nest.map(self._clip_if_enabled,
controller_state)
access_output, access_state = self._access(controller_output,
prev_access_state)
output = tf.concat([controller_output,
batch_flatten(access_output)], 1)
output = snt.Linear(output_size=self._output_size.as_list()[0],
name='output_linear')(output)
output = self._clip_if_enabled(output)
return output, DNCState(access_output=access_output,
access_state=access_state,
controller_state=controller_state)
def __init__(
self,
learning_rate,
num_layers,
size,
size_layer,
output_size,
forget_bias = 0.1,
lambda_coeff = 0.5
):
def lstm_cell(size_layer):
return tf.nn.rnn_cell.GRUCell(size_layer)
rnn_cells = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(size_layer) for _ in range(num_layers)],
state_is_tuple = False,
)
self.X = tf.placeholder(tf.float32, (None, None, size))
self.Y = tf.placeholder(tf.float32, (None, output_size))
drop = tf.contrib.rnn.DropoutWrapper(
rnn_cells, output_keep_prob = forget_bias
)
self.hidden_layer = tf.placeholder(
tf.float32, (None, num_layers * size_layer)
)
_, last_state = tf.nn.dynamic_rnn(
drop, self.X, initial_state = self.hidden_layer, dtype = tf.float32
)
self.z_mean = tf.layers.dense(last_state, size)
self.z_log_sigma = tf.layers.dense(last_state, size)
epsilon = tf.random_normal(tf.shape(self.z_log_sigma))
self.z_vector = self.z_mean + tf.exp(self.z_log_sigma)
with tf.variable_scope('decoder', reuse = False):
rnn_cells_dec = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(size_layer) for _ in range(num_layers)], state_is_tuple = False
)
drop_dec = tf.contrib.rnn.DropoutWrapper(
rnn_cells_dec, output_keep_prob = forget_bias
)
x = tf.concat([tf.expand_dims(self.z_vector, axis=0), self.X], axis = 1)
self.outputs, self.last_state = tf.nn.dynamic_rnn(
drop_dec, self.X, initial_state = last_state, dtype = tf.float32
)
self.logits = tf.layers.dense(self.outputs[-1], output_size)
self.lambda_coeff = lambda_coeff
self.kl_loss = -0.5 * tf.reduce_sum(1.0 + 2 * self.z_log_sigma - self.z_mean ** 2 -
tf.exp(2 * self.z_log_sigma), 1)
self.kl_loss = tf.scalar_mul(self.lambda_coeff, self.kl_loss)
self.cost = tf.reduce_mean(tf.square(self.Y - self.logits) + self.kl_loss)
self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(
self.cost
)
def multihead_attn(queries, keys, q_masks, k_masks, future_binding, num_units,
num_heads):
T_q = tf.shape(queries)[1]
T_k = tf.shape(keys)[1]
Q = tf.layers.dense(queries, num_units, name='Q')
K_V = tf.layers.dense(keys, 2 * num_units, name='K_V')
K, V = tf.split(K_V, 2, -1)
Q_ = tf.concat(tf.split(Q, num_heads, axis=2), axis=0)
K_ = tf.concat(tf.split(K, num_heads, axis=2), axis=0)
V_ = tf.concat(tf.split(V, num_heads, axis=2), axis=0)
align = tf.matmul(Q_, tf.transpose(K_, [0, 2, 1]))
align = align / np.sqrt(K_.get_shape().as_list()[-1])
paddings = tf.fill(tf.shape(align), float('-inf'))
key_masks = k_masks
key_masks = tf.tile(key_masks, [num_heads, 1])
key_masks = tf.tile(tf.expand_dims(key_masks, 1), [1, T_q, 1])
align = tf.where(tf.equal(key_masks, 0), paddings, align)
if future_binding:
lower_tri = tf.ones([T_q, T_k])
lower_tri = tf.linalg.LinearOperatorLowerTriangular(
lower_tri).to_dense()
masks = tf.tile(tf.expand_dims(lower_tri, 0),
[tf.shape(align)[0], 1, 1])
align = tf.where(tf.equal(masks, 0), paddings, align)
align = tf.nn.softmax(align)
query_masks = tf.to_float(q_masks)
query_masks = tf.tile(query_masks, [num_heads, 1])
query_masks = tf.tile(tf.expand_dims(query_masks, -1), [1, 1, T_k])
align *= query_masks
outputs = tf.matmul(align, V_)
outputs = tf.concat(tf.split(outputs, num_heads, axis=0), axis=2)
outputs += queries
outputs = layer_norm(outputs)
return outputs
def sinusoidal_position_encoding(inputs, mask, repr_dim):
T = tf.shape(inputs)[1]
pos = tf.reshape(tf.range(0.0, tf.to_float(T), dtype=tf.float32), [-1, 1])
i = np.arange(0, repr_dim, 2, np.float32)
denom = np.reshape(np.power(10000.0, i / repr_dim), [1, -1])
enc = tf.expand_dims(
tf.concat(
[tf.sin(pos / denom), tf.cos(pos / denom)], 1), 0)
return tf.tile(enc, [tf.shape(inputs)[0], 1, 1]) * tf.expand_dims(
tf.to_float(mask), -1)
def cnn_block(x, dilation_rate, pad_sz, hidden_dim, kernel_size):
x = layer_norm(x)
pad = tf.zeros([tf.shape(x)[0], pad_sz, hidden_dim])
x = tf.layers.conv1d(inputs=tf.concat([pad, x, pad], 1),
filters=hidden_dim,
kernel_size=kernel_size,
dilation_rate=dilation_rate)
x = x[:, :-pad_sz, :]
x = tf.nn.relu(x)
return x
def position_encoding(inputs):
T = tf.shape(inputs)[1]
repr_dim = inputs.get_shape()[-1].value
pos = tf.reshape(tf.range(0.0, tf.to_float(T), dtype=tf.float32), [-1, 1])
i = np.arange(0, repr_dim, 2, np.float32)
denom = np.reshape(np.power(10000.0, i / repr_dim), [1, -1])
enc = tf.expand_dims(
tf.concat(
[tf.sin(pos / denom), tf.cos(pos / denom)], 1), 0)
return tf.tile(enc, [tf.shape(inputs)[0], 1, 1])
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,
forget_bias = 0.1,
):
def lstm_cell(size_layer):
return tf.nn.rnn_cell.LSTMCell(size_layer, state_is_tuple = False)
backward_rnn_cells = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(size_layer) for _ in range(num_layers)],
state_is_tuple = False,
)
forward_rnn_cells = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(size_layer) for _ in range(num_layers)],
state_is_tuple = False,
)
self.X = tf.placeholder(tf.float32, (None, None, size))
self.Y = tf.placeholder(tf.float32, (None, output_size))
drop_backward = tf.contrib.rnn.DropoutWrapper(
backward_rnn_cells, output_keep_prob = forget_bias
)
forward_backward = tf.contrib.rnn.DropoutWrapper(
forward_rnn_cells, output_keep_prob = forget_bias
)
self.backward_hidden_layer = tf.placeholder(
tf.float32, shape = (None, num_layers * 2 * size_layer)
)
self.forward_hidden_layer = tf.placeholder(
tf.float32, shape = (None, num_layers * 2 * size_layer)
)
self.outputs, self.last_state = tf.nn.bidirectional_dynamic_rnn(
forward_backward,
drop_backward,
self.X,
initial_state_fw = self.forward_hidden_layer,
initial_state_bw = self.backward_hidden_layer,
dtype = tf.float32,
)
self.outputs = tf.concat(self.outputs, 2)
self.logits = tf.layers.dense(self.outputs[-1], 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,
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())
