def decoder(input, quan_scale):
with tf.variable_scope('reverse_sigmoid_scale'):
# output = basic_block.reverse_sigmoid(input / 255.0)
# stable reverse sigmoid
output = basic_block.reverse_sigmoid(
(input + 1e-6) / (quan_scale - 1 + 1e-5))
tf.summary.histogram('reverse_sigmoid_scale_result', output)
# output = tf.Print(output, [output[5, 20:30, 20, 0]])
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.identity,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_4')
output = basic_block.res_block(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_res_1',
layer_num=2)
output = basic_block.res_block(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_res_2',
layer_num=2)
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=64,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_3')
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_2')
# output = tf.Print(output, [output[5, 20:30, 20, 0]])
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_1')
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=3,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.identity,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='decode_0')
# output = tf.Print(output, [output[5, 20:30, 20, 0]])
with tf.variable_scope('denormalize'):
output = output * train_data_std + train_data_mean
tf.summary.histogram('denormalized_input', output)
tf.summary.image("recons_image", output, max_outputs=4)
# To be modified
# output = tf.clip_by_value(output, 0.0, 1.0)
output = tf.clip_by_value(output, 0.0, 255.0)
# output = tf.Print(output, [output[5, 20:30, 20, 0]])
return output
def encoder(input, patch_size, quan_scale):
# input = tf.Print(input, [tf.shape(input)[0], tf.shape(input)[1], tf.shape(input)[2], tf.shape(input)[3]], 'Input tensor shape:')
# input = tf.Print(input, [input], 'Input tensor:')
with tf.variable_scope('normalize'):
output = tf.reshape(input, [-1, patch_size, patch_size, 3])
tf.summary.image("ori_image", output, max_outputs=4)
tf.summary.histogram('ori_input', output)
output = (output - train_data_mean) / train_data_std
tf.summary.histogram('normalized_input', output)
# output = tf.Print(output, [output], 'Normalized tensor:')
output = basic_block.my_conv2d(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_0')
output = basic_block.my_conv2d(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_1')
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_2')
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_3')
output = basic_block.res_block(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_res_1',
layer_num=2)
output = basic_block.res_block(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_res_2',
layer_num=2)
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.identity,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='encode_4')
with tf.variable_scope('sigmoid_scale_quantize'):
output = tf.nn.sigmoid(output) * (quan_scale - 1)
output = tf.stop_gradient(tf.round(output) - output) + output
tf.add_to_collection('bitrate_reg_var', output)
tf.summary.histogram('sigmoid_scale_quantize_result', output)
return output
def model(input):
with tf.variable_scope('normalize'):
output = tf.reshape(input, [-1, patch_size, patch_size, 3])
output = (output - train_data_mean) / train_data_std
output = basic_block.my_conv2d(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv_1')
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv_2')
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv_3')
output = basic_block.my_conv2d(
inputs=output,
filters=64,
strides=[1, 1],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv_4')
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=32,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.nn.relu,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv_5')
output = basic_block.my_conv2d_transpose(
inputs=output,
filters=3,
strides=[2, 2],
kernel_size=[3, 3],
padding="SAME",
activation=tf.identity,
kernel_initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01),
bias_initializer=tf.constant_initializer(0.0),
name='conv6')
with tf.variable_scope('denormalize'):
output = output * train_data_std + train_data_mean
output = tf.clip_by_value(output, 0.0, 255.0)
return output
