def layer_6_conv():
from dlnn.layer.Conv2D import Conv2D
from dlnn.layer.Conv2D import AvgFilter
from dlnn.layer.Conv2D import MaxFilter
from dlnn.layer.Conv2D import StdDevFilter
return Conv2D(name='step_6',
filters=[AvgFilter(),
MaxFilter(),
StdDevFilter()],
window=3,
padding='same',
use_bias=False,
kernel_size=(2, 2),
data_format='channels_first',
input_shape=(3, 2, 2))
def build_model(self):
from keras import Input
from keras.activations import sigmoid
from keras.initializers import Zeros
from keras.layers import Reshape, Activation, MaxPooling2D, Flatten, Dense, Concatenate
self.layer['input'] = Input(shape=(self.input_shape, ))
self.layer['pre_scaling'] = Scaling(scale_min1=self.scale_min1,
scale_max1=self.scale_max1,
scale_min2=self.scale_min2,
scale_max2=self.scale_max2,
name='pre_scaling')(
self.layer['input'])
self.layer['pre_reshaping'] = Reshape(
target_shape=[1, 1, self.input_shape],
name='pre_reshaping')(self.layer['pre_scaling'])
self.layer['pre_tiling'] = Tiling(units=self.input_shape,
name='pre_tiling')(
self.layer['pre_reshaping'])
self.layer['cnn_conv_1'] = Conv2D(
filters=[AvgFilter(), MaxFilter(),
StdDevFilter()],
window=self.conv_1_window,
padding='same',
use_bias=False,
kernel_size=(self.input_shape, self.input_shape),
data_format='channels_first',
name='cnn_conv_1')(self.layer['pre_tiling'])
self.layer['cnn_activation_1'] = Activation(
activation=sigmoid,
name='cnn_activation_1')(self.layer['cnn_conv_1'])
self.layer['cnn_conv_2'] = Conv2D(
filters=[AvgFilter(), MaxFilter(),
StdDevFilter()],
window=self.conv_2_window,
padding='same',
use_bias=False,
kernel_size=(self.input_shape, self.input_shape),
data_format='channels_first',
name='cnn_conv_2')(self.layer['cnn_activation_1'])
self.layer['cnn_activation_2'] = Activation(
activation=sigmoid,
name='cnn_activation_2')(self.layer['cnn_conv_2'])
self.layer['cnn_pooling_1'] = MaxPooling2D(
pool_size=(self.pool_1_size, self.pool_1_size),
padding='same',
data_format='channels_first',
name='cnn_pooling_1')(self.layer['cnn_activation_2'])
self.layer['cnn_conv_3'] = Conv2D(
filters=[AvgFilter(), MaxFilter(),
StdDevFilter()],
window=self.conv_3_window,
padding='same',
use_bias=False,
kernel_size=(self.input_shape, self.input_shape),
data_format='channels_first',
name='cnn_conv_3')(self.layer['cnn_pooling_1'])
self.layer['cnn_activation_3'] = Activation(
activation=sigmoid,
name='cnn_activation_3')(self.layer['cnn_conv_3'])
self.layer['cnn_pooling_2'] = MaxPooling2D(
pool_size=(self.pool_2_size, self.pool_2_size),
padding='same',
data_format='channels_first',
name='cnn_pooling_2')(self.layer['cnn_activation_3'])
self.layer['bridge_flatten'] = Flatten(name='bridge_flatten')(
self.layer['cnn_pooling_2'])
self.layer['elm_1_dense_1'] = Dense(
units=self.elm_1_dense_1_units,
activation=None,
use_bias=True,
kernel_initializer=CustomInitializer(corr_step_10_a_kernel_init),
bias_initializer=CustomInitializer(corr_step_10_a_bias_init),
trainable=False,
name='elm_1_dense_1')(self.layer['bridge_flatten'])
self.layer['elm_1_activation_1'] = Activation(
activation=sigmoid,
name='elm_1_activation_1')(self.layer['elm_1_dense_1'])
self.layer['elm_1_dense_2'] = Dense(
units=self.category_num,
activation=None,
use_bias=False,
kernel_initializer=Zeros(),
trainable=False,
name='elm_1_dense_2')(self.layer['elm_1_activation_1'])
self.layer['elm_2_dense_1'] = Dense(
units=self.elm_2_dense_1_units,
activation=None,
use_bias=True,
kernel_initializer=CustomInitializer(corr_step_10_b_kernel_init),
bias_initializer=CustomInitializer(corr_step_10_b_bias_init),
trainable=False,
name='elm_2_dense_1')(self.layer['bridge_flatten'])
self.layer['elm_2_activation_1'] = Activation(
activation=sigmoid,
name='elm_2_activation_1')(self.layer['elm_2_dense_1'])
self.layer['elm_2_dense_2'] = Dense(
units=self.category_num,
activation=None,
use_bias=False,
kernel_initializer=Zeros(),
trainable=False,
name='elm_2_dense_2')(self.layer['elm_2_activation_1'])
self.layer['elm_3_dense_1'] = Dense(
units=self.elm_3_dense_1_units,
activation=None,
use_bias=True,
kernel_initializer=CustomInitializer(corr_step_10_c_kernel_init),
bias_initializer=CustomInitializer(corr_step_10_c_bias_init),
trainable=False,
name='elm_3_dense_1')(self.layer['bridge_flatten'])
self.layer['elm_3_activation_1'] = Activation(
activation=sigmoid,
name='elm_3_activation_1')(self.layer['elm_3_dense_1'])
self.layer['elm_3_dense_2'] = Dense(
units=self.category_num,
activation=None,
use_bias=False,
kernel_initializer=Zeros(),
trainable=False,
name='elm_3_dense_2')(self.layer['elm_3_activation_1'])
self.layer['fully_connected_concat'] = Concatenate(
name='fully_connected_concat')([
self.layer['elm_1_dense_2'], self.layer['elm_2_dense_2'],
self.layer['elm_3_dense_2']
])
self.layer['fully_connected_reshape'] = Reshape(
target_shape=(self.fully_connected_num, self.category_num),
name='fully_connected_reshape')(
self.layer['fully_connected_concat'])
self.layer['fully_connected_merge'] = MergeCategorical(
categorical_length=self.category_num,
name='fully_connected_merge')(
self.layer['fully_connected_reshape'])