def patch_generator(x, is_training):
# input data shape
# 256*256*(3:origin image )
print('-' * 50, 'Input Node Name', '-' * 50)
print(x.name)
print('-' * 120)
k_size = 3
p_size = 2
init_filter = 16
model_parms = dict(pool_size=[p_size, p_size],
kernel_size=k_size,
conv='separable_conv2d',
norm='batch_norm',
activation='relu')
x = tf.reshape(x, [-1] + config_l.input_shape[:], name='Input.x')
block_0 = _bk.conv_norm_activation(x, filters=init_filter, **model_parms)
block_1 = _bk.res_v1_block(block_0, **model_parms)
block_2 = _bk.res_v1_block(block_1, **model_parms)
block_3 = _bk.res_v1_block(block_2, **model_parms)
block_4 = _bk.res_v1_block(block_3, **model_parms)
unet_conv_t_4 = _bk.unet_conv2d_T_block_v0(
block_4,
block_3,
img_size=block_3.shape[1:3].as_list(),
filters=block_3.shape[3].value,
**model_parms)
unet_conv_t_3 = _bk.unet_conv2d_T_block_v0(
unet_conv_t_4,
block_2,
img_size=block_2.shape[1:3].as_list(),
filters=block_2.shape[3].value,
**model_parms)
unet_conv_t_2 = _bk.unet_conv2d_T_block_v0(
unet_conv_t_3,
block_1,
img_size=block_1.shape[1:3].as_list(),
filters=block_1.shape[3].value,
**model_parms)
unet_conv_t_1 = _bk.unet_conv2d_T_block_v0(
unet_conv_t_2,
block_0,
img_size=block_0.shape[1:3].as_list(),
filters=block_0.shape[3].value,
**model_parms)
unet_conv_t_0 = _bk.unet_conv2d_T_block_v0(unet_conv_t_1,
x,
img_size=x.shape[1:3].as_list(),
filters=init_filter)
result = _bk.conv_norm_activation(unet_conv_t_0,
filters=config_l.layer_num,
kernel_size=(1, 1),
**model_parms)
print('-' * 50, 'Result Node Name', '-' * 50)
print(result.name)
print('-' * 120)
return tf.concat([x, result], concat_dim=3)
def patch_segmentation(x, is_training):
# input data shape
# 256*256*(3:origin image )
print('-' * 50, 'Input Node Name', '-' * 50)
print(x.name)
print('-' * 120)
k_size = 3
init_filter = 32
# model_parms=dict(kernel_size=(k_size,k_size),conv=_l.Conv.conv2d,norm=_l.Norm.batch,activation=_op.Activation.leak_relu)
model_parms = dict(kernel_size=(k_size, k_size),
conv=_l.Conv.conv2d,
norm=_l.Norm.batch,
activation=_op.Activation.relu)
block_0 = _bk.conv_norm_activation(x, filters=init_filter, **model_parms)
block_1 = _bk.res_v1_block(block_0, **model_parms)
block_2 = _bk.res_v1_block(block_1, **model_parms)
block_3 = _bk.res_v1_block(block_2, **model_parms)
block_4 = _bk.res_v1_block(block_3, **model_parms)
unet_conv_t_4 = _bk.ushape_block(block_4,
block_3,
img_size=block_3.shape[1:3].as_list(),
filters=block_3.shape[3].value,
**model_parms)
unet_conv_t_3 = _bk.ushape_block(unet_conv_t_4,
block_2,
img_size=block_2.shape[1:3].as_list(),
filters=block_2.shape[3].value,
**model_parms)
unet_conv_t_2 = _bk.ushape_block(unet_conv_t_3,
block_1,
img_size=block_1.shape[1:3].as_list(),
filters=block_1.shape[3].value,
**model_parms)
unet_conv_t_1 = _bk.ushape_block(unet_conv_t_2,
block_0,
img_size=block_0.shape[1:3].as_list(),
filters=block_0.shape[3].value,
**model_parms)
unet_conv_t_0 = _bk.ushape_block(unet_conv_t_1,
x,
img_size=x.shape[1:3].as_list(),
filters=init_filter)
# result=_bk.conv_norm_activation(unet_conv_t_0,filters=_c_s.layer_num,kernel_size=(1,1),activation=None)
result = _bk.conv_norm_activation(unet_conv_t_0,
filters=_c_s.layer_num,
kernel_size=(1, 1),
activation=_op.Activation.sigmoid)
print('-' * 50, 'Result Node Name', '-' * 50)
print(result.name)
print('-' * 120)
return result
def patch_discriminator(x, is_training):
# input data shape
# 256*256*(3:origin image 1:MA heatmap)
print('-' * 50, 'Input Node Name', '-' * 50)
print(x.name)
print('-' * 120)
k_size = 3
p_size = 2
init_filter = 24
model_parms = dict(pool_size=[p_size, p_size],
kernel_size=k_size,
conv='separable_conv2d',
norm='batch_norm',
activation='relu')
x = tf.reshape(x, [-1] + config_l.input_shape[:], name='Input.x')
block_0 = _bk.conv_norm_activation(x, filters=init_filter, **model_parms)
# tf
block_1 = _bk.res_v1_block(block_0, **model_parms)
block_2 = _bk.res_v1_block(block_1, **model_parms)
block_3 = _bk.res_v1_block(block_2, **model_parms)
block_4 = _bk.res_v1_block(block_3, **model_parms)
block_5 = _bk.res_v1_block(block_4, **model_parms)
block_6 = _bk.res_v1_block(block_5, **model_parms)
_y = _l.conv(block_6, _l.Conv.conv2d, filters=2, kernel_size=(1, 1))
_y = _l.pooling(_y, pooling=_l.Pooling.global_avg_2d)
result = _op.activation(result, activation=_op.Activation.softmax)
print('-' * 50, 'Result Node Name', '-' * 50)
print(result.name)
print('-' * 120)
return result
def patch_classification(x, is_training):
# input data shape
# 64*64*(3:origin image )
print('-' * 50, 'Input Node Name', '-' * 50)
print(x.name)
print('-' * 120)
k_size = 3
init_filter = 32
model_parms = dict(kernel_size=(k_size, k_size),
conv=_l.Conv.separable_conv2d,
norm=_l.Norm.batch,
activation=_op.Activation.relu)
# model_parms=dict(kernel_size=(k_size,k_size),conv=_l.Conv.conv2d,norm=_l.Norm.batch,activation=None )
block_0 = _bk.conv_norm_activation(x, filters=init_filter, **model_parms)
block_1 = _bk.res_v1_block(block_0, **model_parms)
block_2 = _bk.res_v1_block(block_1, **model_parms)
block_3 = _bk.res_v1_block(block_2, **model_parms)
block_4 = _bk.res_v1_block(block_3, **model_parms)
class_layer = _l.pooling(block_4, _l.Pooling.global_avg_2d)
class_result = _l.dense(class_layer,
units=_c_s.layer_num,
activation=_op.Activation.relu)
return dict(class_result=class_result,
encode_layers=[block_0, block_1, block_2, block_3, block_4])
def patch_multi_task(x, is_training):
# input data shape
# 256*256*(3:origin image )
print('-' * 50, 'Input Node Name', '-' * 50)
print(x.name)
print('-' * 120)
k_size = 3
init_filter = 32
model_parms = dict(kernel_size=(k_size, k_size),
conv=_l.Conv.separable_conv2d,
norm=_l.Norm.batch,
activation=_op.Activation.relu)
# model_parms=dict(kernel_size=(k_size,k_size),conv=_l.Conv.conv2d,norm=_l.Norm.batch,activation=None )
block_0 = _bk.conv_norm_activation(x, filters=init_filter, **model_parms)
block_1 = _bk.res_v1_block(block_0, **model_parms)
block_2 = _bk.res_v1_block(block_1, **model_parms)
block_3 = _bk.res_v1_block(block_2, **model_parms)
block_4 = _bk.res_v1_block(block_3, **model_parms)
class_layer = _l.pooling(block_4, _l.Pooling.global_avg_2d)
class_result = _l.dense(class_layer,
units=_c_s.layer_num,
activation=_op.Activation.relu)
unet_conv_t_4 = _bk.ushape_block(block_4,
block_3,
img_size=block_3.shape[1:3].as_list(),
filters=block_3.shape[3].value,
**model_parms)
unet_conv_t_3 = _bk.ushape_block(unet_conv_t_4,
block_2,
img_size=block_2.shape[1:3].as_list(),
filters=block_2.shape[3].value,
**model_parms)
unet_conv_t_2 = _bk.ushape_block(unet_conv_t_3,
block_1,
img_size=block_1.shape[1:3].as_list(),
filters=block_1.shape[3].value,
**model_parms)
unet_conv_t_1 = _bk.ushape_block(unet_conv_t_2,
block_0,
img_size=block_0.shape[1:3].as_list(),
filters=block_0.shape[3].value,
**model_parms)
unet_conv_t_0 = _bk.ushape_block(unet_conv_t_1,
x,
img_size=x.shape[1:3].as_list(),
filters=init_filter)
# result=_bk.conv_norm_activation(unet_conv_t_0,filters=_c_s.layer_num,kernel_size=(1,1),activation=None)
seq_result = _bk.conv_norm_activation(unet_conv_t_0,
filters=_c_s.layer_num,
kernel_size=(1, 1),
activation=_op.Activation.relu)
print('-' * 50, 'Result Node Name', '-' * 50)
print(seq_result.name)
print('-' * 120)
return dict(seq_result=seq_result,
class_result=class_result,
encode_layers=[block_0, block_1, block_2, block_3, block_4],
decoder_layers=[
unet_conv_t_4, unet_conv_t_3, unet_conv_t_2, unet_conv_t_1,
unet_conv_t_0
])