def dilated_conv1d(inputs,
filters,
filter_width=2,
dilation_rate=1,
pad='VALID',
activation=None,
name='dialted_conv_1d',
reuse=False):
with tf.name_scope(name):
if dilation_rate > 1:
transformed = time_to_batch(inputs, dilation_rate)
conv = conv1d(x=transformed,
num_filters=filters,
filter_size=filter_width,
stride=1,
pad=pad,
name=name,
reuse=reuse)
restored = batch_to_time(conv, dilation_rate)
else:
restored = conv1d(x=inputs,
num_filters=filters,
filter_size=filter_width,
stride=1,
pad=pad,
name=name,
reuse=reuse)
# Remove excess elements at the end.
out_width = tf.shape(inputs)[1] - (filter_width - 1) * dilation_rate
result = tf.slice(restored, [0, 0, 0], [-1, out_width, -1])
if activation is not None:
result = activation(result)
return result
def conv_1d_network(x,
ob_space,
ac_space,
conv_1d_num_layers=4,
conv_1d_num_filters=32,
conv_1d_filter_size=3,
conv_1d_stride=2,
pad="SAME",
dtype=tf.float32,
collections=None,
reuse=False,
**kwargs):
"""
Stage1 network: from preprocessed 1D input to estimated features.
Encapsulates convolutions, [possibly] skip-connections etc. Can be shared.
Returns:
tensor holding state features;
"""
for i in range(conv_1d_num_layers):
x = tf.nn.elu(
conv1d(
x,
conv_1d_num_filters,
"conv1d_{}".format(i + 1),
conv_1d_filter_size,
conv_1d_stride,
pad,
dtype,
collections,
reuse
)
)
return x
def conv_1d_casual_attention_encoder(
x,
keep_prob,
ob_space=None,
ac_space=None,
conv_1d_num_filters=32,
conv_1d_filter_size=2,
conv_1d_activation=tf.nn.elu,
conv_1d_attention_ref=tf.contrib.seq2seq.LuongAttention,
# conv_1d_attention_ref=tf.contrib.seq2seq.BahdanauAttention,
name='casual_encoder',
reuse=False,
collections=None,
**kwargs):
"""
Tree-shaped convolution stack encoder with self-attention.
Stage1 casual convolutions network: from 1D input to estimated features.
Returns:
tensor holding state features;
"""
with tf.variable_scope(name_or_scope=name, reuse=reuse):
shape = x.get_shape().as_list()
if len(shape) > 3: # remove pseudo 2d dimension
x = x[:, :, 0, :]
num_layers = int(math.log(shape[1], conv_1d_filter_size))
# print('num_layers: ', num_layers)
layers = []
attention_layers = []
y = x
for i in range(num_layers):
_, length, channels = y.get_shape().as_list()
# t2b:
tail = length % conv_1d_filter_size
if tail != 0:
pad = conv_1d_filter_size - tail
paddings = [[0, 0], [pad, 0], [0, 0]]
y = tf.pad(y, paddings)
length += pad
# print('padded_length: ', length)
num_time_batches = int(length / conv_1d_filter_size)
# print('num_time_batches: ', num_time_batches)
y = tf.reshape(y, [-1, conv_1d_filter_size, channels],
name='layer_{}_t2b'.format(i))
y = conv1d(x=y,
num_filters=conv_1d_num_filters,
filter_size=conv_1d_filter_size,
stride=1,
pad='VALID',
name='conv1d_layer_{}'.format(i))
y = tf.nn.dropout(y, keep_prob=keep_prob)
# b2t:
y = tf.reshape(y, [-1, num_time_batches, conv_1d_num_filters],
name='layer_{}_output'.format(i))
y = norm_layer(y)
if conv_1d_activation is not None:
y = conv_1d_activation(y)
layers.append(y)
# Insert attention for all but top layer:
if num_time_batches > 1:
attention = attention_layer(
y,
attention_ref=conv_1d_attention_ref,
name='attention_layer_{}'.format(i))
attention_layers.append(attention)
convolved = tf.stack([h[:, -1, :] for h in layers],
axis=1,
name='convolved_stack')
attended = tf.concat(attention_layers, axis=-2, name='attention_stack')
encoded = tf.concat([convolved, attended],
axis=-2,
name='encoded_state')
# print('convolved: ', convolved)
# print('attention_stack: ', attended)
# print('encoded :', encoded)
return encoded
def conv_1d_casual_encoder(
x,
ob_space,
ac_space,
conv_1d_num_filters=32,
conv_1d_filter_size=2,
conv_1d_activation=tf.nn.elu,
conv_1d_overlap=1,
name='casual_encoder',
keep_prob=None,
conv_1d_gated=False,
reuse=False,
collections=None,
**kwargs
):
"""
Tree-shaped convolution stack encoder as more comp. efficient alternative to dilated one.
Stage1 casual convolutions network: from 1D input to estimated features.
Returns:
tensor holding state features;
"""
with tf.variable_scope(name_or_scope=name, reuse=reuse):
shape = x.get_shape().as_list()
if len(shape) > 3: # remove pseudo 2d dimension
x = x[:, :, 0, :]
num_layers = int(math.log(shape[1], conv_1d_filter_size))
# print('num_layers: ', num_layers)
layers = []
slice_depth = []
y = x
for i in range(num_layers):
_, length, channels = y.get_shape().as_list()
# t2b:
tail = length % conv_1d_filter_size
if tail != 0:
pad = conv_1d_filter_size - tail
paddings = [[0, 0], [pad, 0], [0, 0]]
y = tf.pad(y, paddings)
length += pad
# print('padded_length: ', length)
num_time_batches = int(length / conv_1d_filter_size)
# print('num_time_batches: ', num_time_batches)
y = tf.reshape(y, [-1, conv_1d_filter_size, channels], name='layer_{}_t2b'.format(i))
y = conv1d(
x=y,
num_filters=conv_1d_num_filters,
filter_size=conv_1d_filter_size,
stride=1,
pad='VALID',
name='conv1d_layer_{}'.format(i)
)
# b2t:
y = tf.reshape(y, [-1, num_time_batches, conv_1d_num_filters], name='layer_{}_output'.format(i))
y = norm_layer(y)
if conv_1d_activation is not None:
y = conv_1d_activation(y)
if keep_prob is not None:
y = tf.nn.dropout(y, keep_prob=keep_prob, name="_layer_{}_with_dropout".format(i))
layers.append(y)
depth = conv_1d_overlap // conv_1d_filter_size ** i
if depth < 1:
depth = 1
slice_depth.append(depth)
# encoded = tf.stack([h[:, -1, :] for h in layers], axis=1, name='encoded_state')
sliced_layers = [
tf.slice(
h,
begin=[0, h.get_shape().as_list()[1] - d, 0],
size=[-1, d, -1]
) for h, d in zip(layers, slice_depth)
]
output_stack = sliced_layers
# But:
if conv_1d_gated:
split_size = int(conv_1d_num_filters / 2)
output_stack = []
for l in sliced_layers:
x1 = l[..., :split_size]
x2 = l[..., split_size:]
y = tf.multiply(
x1,
tf.nn.sigmoid(x2),
name='gated_conv_output'
)
output_stack.append(y)
encoded = tf.concat(output_stack, axis=1, name='encoded_state')
# encoded = tf.concat(
# [
# tf.slice(
# h,
# begin=[0, h.get_shape().as_list()[1] - d, 0],
# size=[-1, d, -1]
# ) for h, d in zip(layers, slice_depth)
# ],
# axis=1,
# name='encoded_state'
# )
print('encoder :', encoded)
return encoded
