from PrognosAIs.Model.Architectures.Architecture import ClassificationNetworkArchitecture
from tensorflow.keras.layers import Conv2D
from tensorflow.keras.layers import Conv3D
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import Dropout
from tensorflow.keras.layers import Flatten
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import MaxPooling3D
from tensorflow.keras.models import Model
[docs]class VGG(ClassificationNetworkArchitecture):
[docs] def init_dimensionality(self, N_dimension):
if N_dimension == 2:
self.dims = 2
self.conv_func = Conv2D
self.pool_func = MaxPooling2D
elif N_dimension == 3:
self.dims = 3
self.conv_func = Conv3D
self.pool_func = MaxPooling3D
[docs] def get_VGG_block(self, layer, N_filters, N_conv_layer):
stride_size = self.get_corrected_stride_size(layer, [2] * self.dims, [2] * self.dims)
pooling = self.pool_func(pool_size=[2] * self.dims, strides=stride_size, padding="valid")(
layer
)
conv = self.conv_func(
filters=N_filters,
kernel_size=[3] * self.dims,
strides=[1] * self.dims,
padding="same",
activation="relu",
)(pooling)
for i_conv_layer in range(N_conv_layer - 1):
conv = self.conv_func(
filters=N_filters,
kernel_size=[3] * self.dims,
strides=[1] * self.dims,
padding="same",
activation="relu",
)(conv)
return conv
[docs]class VGG_16_2D(VGG):
dims = 2
[docs] def create_model(self):
self.init_dimensionality(self.dims)
self.check_minimum_input_size(self.inputs, [24, 24])
conv_1 = Conv2D(
filters=64, kernel_size=[3, 3], strides=[1, 1], padding="same", activation="relu",
)(self.inputs)
conv_2 = Conv2D(
filters=64, kernel_size=[3, 3], strides=[1, 1], padding="same", activation="relu",
)(conv_1)
vgg_block_1 = self.get_VGG_block(conv_2, 128, 2)
vgg_block_1 = self.make_dropout_layer(vgg_block_1)
vgg_block_2 = self.get_VGG_block(vgg_block_1, 256, 3)
vgg_block_2 = self.make_dropout_layer(vgg_block_2)
vgg_block_3 = self.get_VGG_block(vgg_block_2, 512, 3)
vgg_block_3 = self.make_dropout_layer(vgg_block_3)
vgg_block_4 = self.get_VGG_block(vgg_block_3, 512, 3)
stride_size = self.get_corrected_stride_size(vgg_block_4, [2, 2], [2, 2])
pooling_1 = MaxPooling2D(pool_size=[2, 2], strides=stride_size, padding="valid",)(
vgg_block_4
)
flatten_1 = Flatten()(pooling_1)
flatten_1 = self.make_dropout_layer(flatten_1)
dense_1 = Dense(4096, activation="relu")(flatten_1)
dense_2 = Dense(4096, activation="relu")(dense_1)
predictions = self.outputs(dense_2)
model = Model(inputs=self.inputs, outputs=predictions)
return model
[docs]class VGG_16_3D(VGG):
dims = 3
[docs] def create_model(self):
self.init_dimensionality(self.dims)
self.check_minimum_input_size(self.inputs, [24, 24, 24])
conv_1 = Conv3D(
filters=64, kernel_size=[3, 3, 3], strides=[1, 1, 1], padding="same", activation="relu",
)(self.inputs)
conv_2 = Conv3D(
filters=64, kernel_size=[3, 3, 3], strides=[1, 1, 1], padding="same", activation="relu",
)(conv_1)
vgg_block_1 = self.get_VGG_block(conv_2, 128, 2)
vgg_block_1 = self.make_dropout_layer(vgg_block_1)
vgg_block_2 = self.get_VGG_block(vgg_block_1, 256, 3)
vgg_block_2 = self.make_dropout_layer(vgg_block_2)
vgg_block_3 = self.get_VGG_block(vgg_block_2, 512, 3)
vgg_block_3 = self.make_dropout_layer(vgg_block_3)
vgg_block_4 = self.get_VGG_block(vgg_block_3, 512, 3)
stride_size = self.get_corrected_stride_size(vgg_block_4, [2, 2, 2], [2, 2, 2])
pooling_1 = MaxPooling3D(pool_size=[2, 2, 2], strides=stride_size, padding="valid",)(
vgg_block_4
)
flatten_1 = Flatten()(pooling_1)
flatten_1 = self.make_dropout_layer(flatten_1)
dense_1 = Dense(4096, activation="relu")(flatten_1)
dense_2 = Dense(4096, activation="relu")(dense_1)
predictions = self.outputs(dense_2)
model = Model(inputs=self.inputs, outputs=predictions)
return model
[docs]class VGG_19_2D(VGG):
dims = 2
[docs] def create_model(self):
self.init_dimensionality(self.dims)
self.check_minimum_input_size(self.inputs, [24, 24])
conv_1 = Conv2D(
filters=64, kernel_size=[3, 3], strides=[1, 1], padding="same", activation="relu",
)(self.inputs)
conv_2 = Conv2D(
filters=64, kernel_size=[3, 3], strides=[1, 1], padding="same", activation="relu",
)(conv_1)
vgg_block_1 = self.get_VGG_block(conv_2, 128, 2)
vgg_block_1 = self.make_dropout_layer(vgg_block_1)
vgg_block_2 = self.get_VGG_block(vgg_block_1, 256, 4)
vgg_block_2 = self.make_dropout_layer(vgg_block_2)
vgg_block_3 = self.get_VGG_block(vgg_block_2, 512, 4)
vgg_block_3 = self.make_dropout_layer(vgg_block_3)
vgg_block_4 = self.get_VGG_block(vgg_block_3, 512, 4)
stride_size = self.get_corrected_stride_size(vgg_block_4, [2, 2], [2, 2])
pooling_1 = MaxPooling2D(pool_size=[2, 2], strides=stride_size, padding="valid",)(
vgg_block_4
)
flatten_1 = Flatten()(pooling_1)
flatten_1 = self.make_dropout_layer(flatten_1)
dense_1 = Dense(4096, activation="relu")(flatten_1)
dense_2 = Dense(4096, activation="relu")(dense_1)
predictions = self.outputs(dense_2)
model = Model(inputs=self.inputs, outputs=predictions)
return model
[docs]class VGG_19_3D(VGG):
dims = 3
[docs] def create_model(self):
self.init_dimensionality(self.dims)
self.check_minimum_input_size(self.inputs, [24, 24, 24])
conv_1 = Conv3D(
filters=64, kernel_size=[3, 3, 3], strides=[1, 1, 1], padding="same", activation="relu",
)(self.inputs)
conv_2 = Conv3D(
filters=64, kernel_size=[3, 3, 3], strides=[1, 1, 1], padding="same", activation="relu",
)(conv_1)
vgg_block_1 = self.get_VGG_block(conv_2, 128, 2)
vgg_block_1 = self.make_dropout_layer(vgg_block_1)
vgg_block_2 = self.get_VGG_block(vgg_block_1, 256, 4)
vgg_block_2 = self.make_dropout_layer(vgg_block_2)
vgg_block_3 = self.get_VGG_block(vgg_block_2, 512, 4)
vgg_block_3 = self.make_dropout_layer(vgg_block_3)
vgg_block_4 = self.get_VGG_block(vgg_block_3, 512, 4)
stride_size = self.get_corrected_stride_size(vgg_block_4, [2, 2, 2], [2, 2, 2])
pooling_1 = MaxPooling3D(pool_size=[2, 2, 2], strides=stride_size, padding="valid",)(
vgg_block_4
)
flatten_1 = Flatten()(pooling_1)
flatten_1 = self.make_dropout_layer(flatten_1)
dense_1 = Dense(4096, activation="relu")(flatten_1)
dense_2 = Dense(4096, activation="relu")(dense_1)
predictions = self.outputs(dense_2)
model = Model(inputs=self.inputs, outputs=predictions)
return model