def conv_lstm_cell_no_input(state,
num_channels,
initializer,
filter_size=5,
forget_bias=1.0,
scope=None,
reuse=None):
with tf.variable_scope(scope,
default_name='BasicConvLstmCell',
values=[state],
reuse=reuse):
state.get_shape().assert_has_rank(4)
c, h = tf.split(3, 2, state)
# Parameters of gates are concatenated into one conv for efficiency.
i_j_f_o = layers.conv2d(h,
4 * num_channels, [filter_size, filter_size],
stride=1,
activation_fn=None,
scope='Gates',
weights_initializer=initializer)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(3, 4, i_j_f_o)
new_c = c * tf.sigmoid(f + forget_bias) + tf.sigmoid(i) * tf.tanh(j)
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(3, [new_c, new_h])
def basic_conv_lstm_cell_leakyrelu_norm(inputs,
state,
num_channels,
filter_size=5,
rate=1,
forget_bias=1.0,
stride=1,
scope=None,
reuse=None):
"""LSTM with leakyRELU activation and layer normalization.
We add forget_bias (default: 1) to the biases of the forget gate in order to
reduce the scale of forgetting in the beginning of the training.
It does not allow cell clipping, a projection layer, and does not
use peep-hole connections: it is the basic baseline.
Args:
inputs: input Tensor, 4D, batch x height x width x channels.
state: state Tensor, 4D, batch x height x width x channels.
num_channels: the number of output channels in the layer.
filter_size: the shape of the each convolution filter.
forget_bias: the initial value of the forget biases.
scope: Optional scope for variable_scope.
reuse: whether or not the layer and the variables should be reused.
Returns:
a tuple of tensors representing output and the new state.
"""
spatial_size = inputs.get_shape()[1:3]
if state is None:
state = init_state(inputs, list(spatial_size) + [2 * num_channels])
with tf.variable_scope(scope,
'BasicConvLstmCell', [inputs, state],
reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state.get_shape().assert_has_rank(4)
c, h = tf.split(axis=3, num_or_size_splits=2, value=state)
inputs_h = tf.concat(axis=3, values=[inputs, h])
# Parameters of gates are concatenated into one conv for efficiency.
i_j_f_o = layers.conv2d(inputs_h,
4 * num_channels, [filter_size, filter_size],
stride=stride,
rate=rate,
activation_fn=None,
scope='Gates')
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(axis=3, num_or_size_splits=4, value=i_j_f_o)
new_c = c * tf.sigmoid(f + forget_bias) + tf.sigmoid(
i) * tf.nn.leaky_relu(j, alpha=0.1)
new_c = tf_layers.layer_norm(new_c, scope="normalisation", reuse=reuse)
new_h = tf.nn.leaky_relu(new_c, alpha=0.1) * tf.sigmoid(o)
return new_h, tf.concat(axis=3, values=[new_c, new_h])
def basic_conv_lstm_cell(inputs,
state,
num_channels,
filter_size=5,
forget_bias=1.0,
scope=None,
reuse=None):
"""Basic LSTM recurrent network cell, with 2D convolution connctions.
具有2D卷积连接的基本LSTM递归网络单元
We add forget_bias (default: 1) to the biases of the forget gate in order to
reduce the scale of forgetting in the beginning of the training.
It does not allow cell clipping, a projection layer, and does not
use peep-hole connections: it is the basic baseline.
为了减少训练开始时的遗忘规模，我们在忘记门的偏差上添加了ignore_bias（默认值：1）。
Args:
inputs: input Tensor, 4D, batch x height x width x channels.
state: state Tensor, 4D, batch x height x width x channels.
num_channels: the number of output channels in the layer.
图层中输出通道的数量
filter_size: the shape of the each convolution filter.
过滤器大小
forget_bias: the initial value of the forget biases.
遗忘偏差
scope: Optional scope for variable_scope.
可选范围
reuse: whether or not the layer and the variables should be reused.
是否重用图层和变量
Returns:
a tuple of tensors representing output and the new state.
代表输出和新状态的张量的元组
"""
spatial_size = inputs.get_shape()[1:3]
if state is None:
state = init_state(inputs, list(spatial_size) + [2 * num_channels])
with tf.variable_scope(scope,
'BasicConvLstmCell', [inputs, state],
reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state.get_shape().assert_has_rank(4)
c, h = tf.split(axis=3, num_or_size_splits=2, value=state)
inputs_h = tf.concat(axis=3, values=[inputs, h])
# Parameters of gates are concatenated into one conv for efficiency.
#为了提高效率，将Gates的参数串联到一个conv中。
i_j_f_o = layers.conv2d(inputs_h,
4 * num_channels, [filter_size, filter_size],
stride=1,
activation_fn=None,
scope='Gates')
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(axis=3, num_or_size_splits=4, value=i_j_f_o)
new_c = c * tf.sigmoid(f + forget_bias) + tf.sigmoid(i) * tf.tanh(j)
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(axis=3, values=[new_c, new_h])
def basic_conv_lstm_cell(inputs,
state,
num_channels,
filter_size=5,
forget_bias=1.0,
scope=None,
reuse=None,
device_for_variables=None):
"""Basic LSTM recurrent network cell, with 2D convolution connctions.
We add forget_bias (default: 1) to the biases of the forget gate in order to
reduce the scale of forgetting in the beginning of the training.
It does not allow cell clipping, a projection layer, and does not
use peep-hole connections: it is the basic baseline.
Args:
inputs: input Tensor, 4D, batch x height x width x channels.
state: state Tensor, 4D, batch x height x width x channels.
num_channels: the number of output channels in the layer.
filter_size: the shape of the each convolution filter.
forget_bias: the initial value of the forget biases.
scope: Optional scope for variable_scope.
reuse: whether or not the layer and the variables should be reused.
Returns:
a tuple of tensors representing output and the new state.
"""
spatial_size = inputs.get_shape()[1:3]
if state is None:
state = init_state(inputs, list(spatial_size) + [2 * num_channels])
with tf.variable_scope(scope,
'BasicConvLstmCell',
[inputs, state],
reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state.get_shape().assert_has_rank(4)
c, h = tf.split(3, 2, state)
inputs_h = tf.concat(3, [inputs, h])
# Parameters of gates are concatenated into one conv for efficiency.
i_j_f_o = layers.conv2d(inputs_h,
4 * num_channels, [filter_size, filter_size],
stride=1,
activation_fn=None,
scope='Gates',
device_for_variables= device_for_variables)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(3, 4, i_j_f_o)
new_c = c * tf.sigmoid(f + forget_bias) + tf.sigmoid(i) * tf.tanh(j)
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(3, [new_c, new_h])
def basic_conv_lstm_cell(inputs,
state,
num_channels,
filter_size=5,
forget_bias=1.0,
scope=None,
reuse=None):
"""Basic LSTM recurrent network cell, with 2D convolution connctions.
We add forget_bias (default: 1) to the biases of the forget gate in order to
reduce the scale of forgetting in the beginning of the training.
It does not allow cell clipping, a projection layer, and does not
use peep-hole connections: it is the basic baseline.
Args:
inputs: input Tensor, 4D, batch x height x width x channels.
state: state Tensor, 4D, batch x height x width x channels.
num_channels: the number of output channels in the layer.
filter_size: the shape of the each convolution filter.
forget_bias: the initial value of the forget biases.
scope: Optional scope for variable_scope.
reuse: whether or not the layer and the variables should be reused.
Returns:
a tuple of tensors representing output and the new state.
"""
spatial_size = inputs.get_shape()[1:3]
if state is None:
state = init_state(inputs, list(spatial_size) + [2 * num_channels])
with tf.variable_scope(scope,
'BasicConvLstmCell',
[inputs, state],
reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state.get_shape().assert_has_rank(4)
c, h = tf.split(axis=3, num_or_size_splits=2, value=state)
inputs_h = tf.concat(axis=3, values=[inputs, h])
# Parameters of gates are concatenated into one conv for efficiency.
i_j_f_o = layers.conv2d(inputs_h,
4 * num_channels, [filter_size, filter_size],
stride=1,
activation_fn=None,
scope='Gates')
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(axis=3, num_or_size_splits=4, value=i_j_f_o)
new_c = c * tf.sigmoid(f + forget_bias) + tf.sigmoid(i) * tf.tanh(j)
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(axis=3, values=[new_c, new_h])
def cubic_lstm_cell(inputs,
state_x,
state_y,
num_channels,
filter_size_x=3,
filter_size_y=1,
filter_size_z=5,
forget_bias=1.0,
scope=None,
reuse=None):
spatial_size = inputs.get_shape()[1:3]
if state_x is None:
state_x = init_state(inputs, list(spatial_size) + [2 * num_channels])
if state_y is None:
state_y = init_state(inputs, list(spatial_size) + [2 * num_channels])
with tf.variable_scope(scope,
'CubicLstmCell', [inputs, state_x, state_y],
reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state_x.get_shape().assert_has_rank(4)
state_y.get_shape().assert_has_rank(4)
c_x, h_x = tf.split(axis=3, num_or_size_splits=2, value=state_x)
c_y, h_y = tf.split(axis=3, num_or_size_splits=2, value=state_y)
inputs_h = tf.concat(axis=3, values=[inputs, h_x, h_y])
# Spatial
i_j_f_o_y = layers.conv2d(inputs_h,
4 * num_channels,
[filter_size_x, filter_size_x],
stride=1,
activation_fn=None,
scope='GatesY')
i_y, j_y, f_y, o_y = tf.split(axis=3,
num_or_size_splits=4,
value=i_j_f_o_y)
new_c_y = c_y * tf.sigmoid(f_y + forget_bias) + tf.sigmoid(
i_y) * tf.tanh(j_y)
new_h_y = tf.tanh(new_c_y) * tf.sigmoid(o_y)
# Temporal
i_j_f_o_x = layers.conv2d(inputs_h,
4 * num_channels,
[filter_size_y, filter_size_y],
stride=1,
activation_fn=None,
scope='GatesX')
i_x, j_x, f_x, o_x = tf.split(axis=3,
num_or_size_splits=4,
value=i_j_f_o_x)
new_c_x = c_x * tf.sigmoid(f_x + forget_bias) + tf.sigmoid(
i_x) * tf.tanh(j_x)
new_h_x = tf.tanh(new_c_x) * tf.sigmoid(o_x)
# Output
new_h = layers.conv2d(tf.concat(axis=3, values=[new_h_x, new_h_y]),
num_channels, [filter_size_z, filter_size_z],
stride=1,
activation_fn=None,
scope='Output')
return new_h, tf.concat(axis=3, values=[new_c_x, new_h_x]), tf.concat(
axis=3, values=[new_c_y, new_h_y])
def TrajGRUCell(inputs,
state,
num_channels,
filter_size=5,
forget_bias=1.0,
scope=None,
reuse=None):
"""Basic LSTM recurrent network cell, with 2D convolution connctions.
We add forget_bias (default: 1) to the biases of the forget gate in order to
reduce the scale of forgetting in the beginning of the training.
It does not allow cell clipping, a projection layer, and does not
use peep-hole connections: it is the basic baseline.
Args:
inputs: input Tensor, 4D, batch x height x width x channels.
state: state Tensor, 4D, batch x height x width x channels.
num_channels: the number of output channels in the layer.
filter_size: the shape of the each convolution filter.
forget_bias: the initial value of the forget biases.
scope: Optional scope for variable_scope.
reuse: whether or not the layer and the variables should be reused.
Returns:
a tuple of tensors representing output and the new state.
"""
spatial_size = inputs.get_shape()[1:3]
ti = tf.range(spatial_size[0])
tj = tf.range(spatial_size[1])
# chn = inputs.get_shape()[3]
mi = tf.meshgrid(ti, tj)[0]
mj = tf.transpose(mi)
mi = tf.cast(tf.reshape(mi, [1, spatial_size[0], spatial_size[1], 1]),
dtype=tf.float32)
mj = tf.cast(tf.reshape(mj, [1, spatial_size[1], spatial_size[0], 1]),
dtype=tf.float32)
if state is None:
state = init_state(inputs, list(spatial_size) + [num_channels])
with tf.variable_scope(scope, 'TrajGRUCell', [inputs, state], reuse=reuse):
inputs.get_shape().assert_has_rank(4)
state.get_shape().assert_has_rank(4)
# ow, h = tf.split(axis=3, num_or_size_splits=2, value=state)
h = state
# h=state
inputs_h = tf.concat(axis=3, values=[inputs, h])
# TrajGRU
u = layers.conv2d(inputs_h, num_channels, [5, 5], scope="u")
v = layers.conv2d(inputs_h, num_channels, [5, 5], scope="v")
z_r_o = layers.conv2d(inputs,
3 * num_channels, [filter_size, filter_size],
stride=1,
activation_fn=None,
scope='Gates')
z, r, o = tf.split(axis=3, num_or_size_splits=3, value=z_r_o)
# ============== 开写
mi = tf.tile(mi, [inputs.get_shape()[0], 1, 1, num_channels])
mj = tf.tile(mj, [inputs.get_shape()[0], 1, 1, num_channels])
wp = warp(h, u, v, mi, mj)
# if not warpfields:
# warpfields = []
# warpfields.append(wp)
# sumw = 0
# for i in range(L):
# tscope = "w%d" % i
# if i < warpfields.__len__():
# tmp = warpfields[-1 * i]
# w = layers.conv2d(tmp, 3 * num_channels, [1, 1], stride=1,
# scope=tscope)
# sumw += w
# else:
# tmp = warpfields[-1]
# w = layers.conv2d(tmp, 3 * num_channels, [1, 1], stride=1,
# scope=tscope)
w = layers.conv2d(wp, 3 * L, [1, 1], stride=1, scope='warp')
zw, rw, ow = tf.split(axis=3, num_or_size_splits=3, value=w)
# w = tf.reduce_sum(w, axis=3)
# sumw += w
# w = tf.reduce_sum(w, axis=0)
zw = tf.reshape(tf.reduce_sum(zw, axis=3), [
inputs.get_shape()[0],
inputs.get_shape()[1],
inputs.get_shape()[2], 1
])
rw = tf.reshape(tf.reduce_sum(rw, axis=3), [
inputs.get_shape()[0],
inputs.get_shape()[1],
inputs.get_shape()[2], 1
])
ow = tf.reshape(tf.reduce_sum(ow, axis=3), [
inputs.get_shape()[0],
inputs.get_shape()[1],
inputs.get_shape()[2], 1
])
z = tf.sigmoid(z + tf.tile(zw, [1, 1, 1, num_channels]))
r = tf.sigmoid(r + tf.tile(rw, [1, 1, 1, num_channels]))
# 论文中f函数,作者没描述,这里用tanh替代
# hp = tf.tanh(o + tf.multiply(r, sumw))
hp = tf.tanh(o + r * tf.tile(ow, [1, 1, 1, num_channels]))
# new_h = tf.multiply(1 - z, hp) + tf.multiply(z, h)
new_h = (1 - z) * hp + z * h
return new_h, new_h