def mnist_feat_ext(main_input, l2_weight=0.0):
net = dnn.Convolution2D(32, (5, 5), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(main_input)
net = dnn.MaxPooling2D(pool_size=(2,2), strides=2)(net)
net = dnn.Convolution2D(48, (5,5),padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2,2),strides=2)(net)
net = dnn.Flatten()(net)
net = dnn.Dense(100,activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Dense(100,activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
return net
def mnist_fcn_featext(main_input, l2_weight=0.0):
shape = main_input.shape.as_list()
if len(shape)==4:
net = dnn.Lambda(lambda x: dnn.keras.backend.reshape(x, (-1, shape[1]*shape[2]*shape[3])))(main_input)
else:
net = main_input
net = dnn.Dense(500,activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
# net = dnn.Dropout(0.2)(net)
net = dnn.Dense(500,activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
return net
def assda_feat_ext(main_input, l2_weight=0.0):
net = dnn.Convolution2D(32, (3, 3),padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(main_input)
net = dnn.Convolution2D(32, (3, 3), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=1)(net)
net = dnn.Convolution2D(64, (3, 3), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Convolution2D(64, (3, 3), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=1)(net)
net = dnn.Convolution2D(128, (3, 3), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Convolution2D(128, (3, 3), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=1)(net)
#
net = dnn.Flatten()(net)
net = dnn.Dense(128,activation='sigmoid',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight),name='feat_ext')(net)
# net = dnn.Dense(512,activation='sigmoid',
# kernel_regularizer=dnn.keras.regularizers.l2(l2_weight),name='feat_ext')(net)
return net
def classifier_dropout(model_input, nclass,l2_weight=0.0):
# net = dnn.Dense(128,activation='relu',
# kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(model_input)
# net = dnn.Dense(128,activation='relu',
# kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Dropout(0.2)(model_input)
net = dnn.Dense(nclass,activation='softmax', name='classifier_output')(net)
return net
def softmax_classification(input_shape, nclass, l2_weight=0):
model = dnn.Sequential()
model.add(
dnn.Dense(nclass,
input_shape=(input_shape, ),
activation='softmax',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
return model
def mnist_reconst(model_input, out_shape):
x = dnn.Dense(out_shape[1]*out_shape[2]*out_shape[3], activation='relu')(model_input)
x = dnn.Reshape((out_shape[1], out_shape[2], out_shape[3]))(x)
x = dnn.Convolution2D(64, (3,3), activation='relu',padding='same')(x)
x = dnn.UpSampling2D((2,2))(x)
x = dnn.Convolution2D(64, (3,3), activation='relu',padding='same')(x)
x = dnn.UpSampling2D((2,2))(x)
x= dnn.Convolution2D(1, (3,3), activation='linear',padding='same', name ='decoder')(x)
return x
def mnist_featext(main_input, out_shape=False):
net = dnn.Convolution2D(32, (3,3), activation='relu', padding='same')(main_input)
net = dnn.MaxPooling2D((2,2),padding='same')(net)
net = dnn.Convolution2D(64, (3,3), activation='relu',padding='same')(net)
x1 = dnn.MaxPooling2D((2,2),padding='same')(net)
net = dnn.Flatten()(x1)
net = dnn.Dense(256, activation='relu', name= 'encoder')(net)
if out_shape:
return net, x1.get_shape().as_list()
else:
return net
def svhnn_model(img_shape, n_class, l2_weight=0):
model = dnn.Sequential()
model.add(dnn.Convolution2D(64, (5, 5), input_shape=(img_shape[1], img_shape[2],img_shape[3]),
padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
model.add(dnn.MaxPooling2D(pool_size=(3,3), strides=2))
# model.add(dnn.Dropout(0.9))
model.add(dnn.Convolution2D(64, (5, 5), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
model.add(dnn.MaxPooling2D(pool_size=(3,3), strides=2))
# model.add(dnn.Dropout(0.75))
model.add(dnn.Convolution2D(128, (5,5), padding='same', activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
# model.add(dnn.Dropout(0.75))
model.add(dnn.Flatten())
model.add(dnn.Dense(3072,activation='relu',kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
# model.add(dnn.Dropout(0.25))
model.add(dnn.Dense(2048,activation='relu',kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
# model.add(dnn.Dropout(0.25))
model.add(dnn.Dense(n_class,activation='softmax',kernel_regularizer=dnn.keras.regularizers.l2(l2_weight)))
return model
def assda_feat_ext(main_input, l2_weight=0.0, small_model=False):
padding = 'same'
maxpool_strides = 2 if small_model else 1
net = dnn.Convolution2D(
32, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(main_input)
net = dnn.Convolution2D(
32, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=maxpool_strides)(net)
# net = dnn.Dropout(0.5)(net)
net = dnn.Convolution2D(
64, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Convolution2D(
64, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=maxpool_strides)(net)
# net = dnn.Dropout(0.5)(net)
net = dnn.Convolution2D(
128, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.Convolution2D(
128, (3, 3),
padding=padding,
activation='relu',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight))(net)
net = dnn.MaxPooling2D(pool_size=(2, 2), strides=maxpool_strides)(net)
#
net = dnn.Flatten()(net)
net = dnn.Dense(128,
activation='sigmoid',
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight),
name='feat_ext')(net)
return net
def cifar10_reconst(model_input, out_shape, l2_weight=0.0, name_prefix = 'sup_'):
x = dnn.Dense(out_shape[1]*out_shape[2]*out_shape[3])(model_input)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Reshape((out_shape[1], out_shape[2], out_shape[3]))(x)
# block 3
x = dnn.Convolution2D(196, (3, 3), padding='same', kernel_initializer="he_normal",
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight),name=name_prefix+'de_block3_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Convolution2D(196, (3, 3), padding='same', kernel_initializer="he_normal",
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight),name=name_prefix+'de_block3_conv1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.UpSampling2D((2,2),name=name_prefix+'block3_pool')(x)
# block 2
x = dnn.Conv2D(128, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'de_block2_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Conv2D(128, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'de_block2_conv1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.UpSampling2D((2,2),name=name_prefix+'block2_pool')(x)
# block 1
x = dnn.Conv2D(64, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'de_block1_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Conv2D(64, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'de_block1_conv1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.UpSampling2D((2, 2),name=name_prefix+'block1_pool')(x)
#
x = dnn.Convolution2D(3, (3, 3), padding='same', kernel_initializer="he_normal",
kernel_regularizer=dnn.keras.regularizers.l2(l2_weight), name =name_prefix+'decoder')(x)
#x= dnn.BatchNormalization()(x)
return x
def cifar10_featext(main_input, l2_weight=0.01, out_shape=False, name_prefix = 'sup_'):
x = dnn.Conv2D(64, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block1_conv1')(main_input)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Conv2D(64, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block1_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.MaxPooling2D((2, 2), strides=(2, 2), name=name_prefix+'block1_pool')(x)
# Block 2
x = dnn.Conv2D(128, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block2_conv1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Conv2D(128, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block2_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.MaxPooling2D((2, 2), strides=(2, 2), name=name_prefix+'block2_pool')(x)
# Block 3
x = dnn.Conv2D(196, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block3_conv1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x = dnn.Conv2D(196, (3, 3), padding='same', kernel_initializer="he_normal", name=name_prefix+'block3_conv2')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu')(x)
x1 = dnn.MaxPooling2D((2, 2), strides=(2, 2), name=name_prefix+'block3_pool')(x)
x = dnn.Flatten(name=name_prefix+'flatten')(x1)
x = dnn.Dense(256, kernel_initializer="he_normal", kernel_regularizer=dnn.l2(l2_weight), bias_regularizer=dnn.l2(l2_weight), name=name_prefix+'fc1')(x)
x = dnn.BatchNormalization()(x)
x = dnn.Activation('relu', name=name_prefix+'encoder')(x)
if out_shape ==False:
return x
else:
return x, x1.get_shape().as_list()
def classifier(model_input, nclass, l2_weight=0.0):
net = dnn.Dense(100, activation='relu', name='cl')(model_input)
net = dnn.Dense(nclass, activation='softmax', name='cl_output')(net)
return net
def feat_ext(main_input, l2_weight=0.0):
net = dnn.Dense(500, activation='relu', name='fe')(main_input)
net = dnn.Dense(100, activation='relu', name='feat_ext')(net)
return net
def labelpredictor(model_input, nclass,l2_weight=0.0, drop_out=None):
if drop_out is not None:
model_input = dnn.Dropout(drop_out)(model_input)
net = dnn.Dense(nclass,activation='softmax', name='classifier_output')(model_input)
return net
def classifier_dropout(model_input, nclass, l2_weight=0.0):
net = dnn.Dropout(0.5)(model_input)
net = dnn.Dense(nclass, activation='softmax',
name='classifier_output')(net)
return net
def regressor(model_input, noutputs, l2_weight=0.0):
net = dnn.Dense(noutputs, activation='sigmoid',
name='regressor_output')(model_input)
return net
