def mask_2D_resiBlock(x, filter_nums):
w = basic_DL_op.weight_variable('conv1', [3, 3, filter_nums, filter_nums],
0.01)
mask = [[1, 1, 1], [1, 1, 0], [0, 0, 0]]
mask = tf.reshape(mask, shape=[3, 3, 1, 1])
mask = tf.tile(mask, multiples=[1, 1, filter_nums, filter_nums])
mask = tf.cast(mask, dtype=tf.float32)
w = w * mask
b = basic_DL_op.bias_variable('bias1', [filter_nums])
c = basic_DL_op.conv2d(x, w) + b
c = tf.nn.relu(c)
w = basic_DL_op.weight_variable('conv2', [3, 3, filter_nums, filter_nums],
0.01)
w = w * mask
b = basic_DL_op.bias_variable('bias2', [filter_nums])
c = basic_DL_op.conv2d(c, w) + b
return x + c
def p_block(x):
w = basic_DL_op.weight_variable('conv1', [3, 3, 1, 16], 0.01)
b = basic_DL_op.bias_variable('bias1', [16])
tmp = basic_DL_op.conv2d_pad(x, w) + b
conv1 = tmp
tmp = tf.nn.tanh(tmp)
w = basic_DL_op.weight_variable('conv2', [3, 3, 16, 16], 0.01)
b = basic_DL_op.bias_variable('bias2', [16])
tmp = basic_DL_op.conv2d_pad(tmp, w) + b
tmp = tf.nn.tanh(tmp)
w = basic_DL_op.weight_variable('conv3', [3, 3, 16, 16], 0.01)
b = basic_DL_op.bias_variable('bias3', [16])
tmp = basic_DL_op.conv2d_pad(tmp, w) + b
tmp = conv1 + tmp
w = basic_DL_op.weight_variable('conv4', [3, 3, 16, 1], 0.01)
b = basic_DL_op.bias_variable('bias4', [1])
out = basic_DL_op.conv2d_pad(tmp, w) + b
return out
def resiBlock_2D_context(x, filter_nums):
w = basic_DL_op.weight_variable('conv1_c',
[3, 3, filter_nums, filter_nums], 0.01)
b = basic_DL_op.bias_variable('bias1_c', [filter_nums])
c = basic_DL_op.conv2d(x, w) + b
c = tf.nn.relu(c)
w = basic_DL_op.weight_variable('conv2_c',
[3, 3, filter_nums, filter_nums], 0.01)
b = basic_DL_op.bias_variable('bias2_c', [filter_nums])
c = basic_DL_op.conv2d(c, w) + b
return x + c
def lstm_layer(x, h, c, in_num, out_num):
# the first layer: input
w = basic_DL_op.weight_variable('conv1', [3, 3, in_num, out_num], 0.01)
x = basic_DL_op.conv2d(x, w)
# the first layer: state
w = basic_DL_op.weight_variable('conv2', [3, 3, out_num, out_num], 0.01)
h = basic_DL_op.conv2d(h, w)
b = basic_DL_op.bias_variable('bias', [out_num])
c, h = lstm_logic(x + h + b, c)
return c, h
def deconv_layer(x):
x_shape = x.get_shape().as_list()
kernel = basic_DL_op.weight_variable('deconv',
[3, 3, x_shape[3], x_shape[3]], 0.01)
x = tf.nn.conv2d_transpose(x,
kernel,
output_shape=[
x_shape[0],
int(x_shape[1] * 2),
int(x_shape[2] * 2), x_shape[3]
],
strides=[1, 2, 2, 1],
padding="SAME")
return x
def mask_2D_layer(x,
static_QP,
context,
features=128,
resi_num=2,
para_num=58):
x = x / static_QP
label = x
# x = tf.stop_gradient(x)
################## layer 1, linear
w = basic_DL_op.weight_variable('conv1', [3, 3, 1, features], 0.01)
w_c = basic_DL_op.weight_variable('conv1_c', [3, 3, 1, features], 0.01)
mask = [[1, 1, 1], [1, 0, 0], [0, 0, 0]]
mask = tf.reshape(mask, shape=[3, 3, 1, 1])
mask = tf.tile(mask, multiples=[1, 1, 1, features])
mask = tf.cast(mask, dtype=tf.float32)
w = w * mask
b = basic_DL_op.bias_variable('bias1', [features])
b_c = basic_DL_op.bias_variable('bias1_c', [features])
x = basic_DL_op.conv2d(x, w) + b
context = basic_DL_op.conv2d(context, w_c) + b_c
conv1 = x
x = x + context
################## layers: resi_num resi_block
for i in range(resi_num):
with tf.variable_scope('resi_block' + str(i)):
x = mask_2D_resiBlock(x, features)
context = resiBlock_2D_context(context, features)
x = x + context
x = conv1 + x
################# conv: after skip connection, relu
w = basic_DL_op.weight_variable('conv2', [3, 3, features, features], 0.01)
mask = [[1, 1, 1], [1, 1, 0], [0, 0, 0]]
mask = tf.reshape(mask, shape=[3, 3, 1, 1])
mask = tf.tile(mask, multiples=[1, 1, features, features])
mask = tf.cast(mask, dtype=tf.float32)
w = w * mask
b = basic_DL_op.bias_variable('bias2', [features])
x = basic_DL_op.conv2d(x, w) + b
x = tf.nn.relu(x)
################# convs: 1x1, relu/linear
w = basic_DL_op.weight_variable('conv3', [1, 1, features, features], 0.01)
b = basic_DL_op.bias_variable('bias3', [features])
x = basic_DL_op.conv2d(x, w) + b
x = tf.nn.relu(x)
w = basic_DL_op.weight_variable('conv4', [1, 1, features, features], 0.01)
b = basic_DL_op.bias_variable('bias4', [features])
x = basic_DL_op.conv2d(x, w) + b
x = tf.nn.relu(x)
w = basic_DL_op.weight_variable('conv5', [1, 1, features, para_num], 0.01)
b = basic_DL_op.bias_variable('bias5', [para_num])
x = basic_DL_op.conv2d(x, w) + b
################# cal the cdf with the output params
h = tf.nn.softplus(x[:, :, :, 0:33])
b = x[:, :, :, 33:46]
a = tf.tanh(x[:, :, :, 46:58])
lower = label - 0.5 / static_QP
high = label + 0.5 / static_QP
lower = cal_cdf(lower, h, b, a)
high = cal_cdf(high, h, b, a)
prob = tf.maximum((high - lower), 1e-9)
cross_entropy = -tf.reduce_mean(tf.log(prob))
return cross_entropy