def init_model(self, input_shape, num_classes): #_alexNet
n_filters = 96
self.model = Sequential()
self.model.add(BatchNormalization(input_shape=input_shape))
self.model.add(
Conv2D(n_filters, (11, 11),
activation='relu',
strides=(2, 2),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2)))
N = self.model.layers[-1].output.get_shape().as_list()[3]
self.model.add(
Conv2D(N, (5, 5),
activation='relu',
strides=(2, 2),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(2, 2)))
N = self.model.layers[-1].output.get_shape().as_list()[3]
self.model.add(
Conv2D(N, (3, 3),
activation='relu',
strides=(1, 1),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(2, 2)))
self.model.add(Flatten())
self.model.add(Dropout(0.5))
self.model.add(
Dense(numpy.prod(
self.model.layers[-3].output.get_shape().as_list()[1:]),
activation='relu'))
self.model.add(Dense(num_classes, activation='softmax'))
weights = tools_CNN_view.import_weight(
'data/filter1.png', (11, 11, 3, 96))[:, :, :, :n_filters]
self.model.layers[1].set_weights(
[weights, tools_CNN_view.import_bias()])
self.model.layers[1].trainable = True
self.model.summary()
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
tools_CNN_view.visualize_filters(self.model, self.folder_debug)
tools_CNN_view.visualize_layers(self.model, self.filename_debug_image,
self.folder_debug)
return
def init_model_2x2(self, input_shape, num_classes):
n_filters = 64
n_features = 1 * num_classes * num_classes
n_features = 64
scale = int(
math.sqrt(n_filters) *
math.sqrt(input_shape[0] * input_shape[1] / n_features))
d1 = int(math.sqrt(scale))
d2 = int(scale / d1)
self.S1 = int(math.sqrt(d1))
self.S2 = int(d1 / self.S1)
self.S3 = int(math.sqrt(d2))
self.S4 = int(d2 / self.S3)
self.model = Sequential()
self.model.add(BatchNormalization(input_shape=input_shape))
self.model.add(
Conv2D(n_filters, (2, 2),
activation='relu',
strides=(self.S1, self.S1),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(self.S2, self.S2)))
N = self.model.layers[-1].output.get_shape().as_list()[3]
self.model.add(
Conv2D(N, (2, 2),
activation='relu',
strides=(self.S3, self.S3),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(self.S4, self.S4)))
self.model.add(Flatten())
self.model.add(Dropout(0.5))
N = numpy.prod(self.model.layers[-3].output.get_shape().as_list()[1:])
self.model.add(Dense(N, activation='relu'))
self.model.add(Dense(num_classes, activation='softmax'))
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
tools_CNN_view.visualize_layers(self.model, self.filename_debug_image,
self.folder_debug)
self.model.summary()
return
def visualize_layers_keras_MobileNet():
filename_image = 'data/ex_natural_images/dog/dog_0101.jpg'
path_out = 'data/output/'
if not os.path.exists(path_out):
os.makedirs(path_out)
else:
tools_IO.remove_files(path_out)
CNN = CNN_App_Keras.CNN_App_Keras()
tools_CNN_view.visualize_filters(CNN.model, path_out)
tools_CNN_view.visualize_layers(CNN.model, filename_image, path_out)
return
def visualize_layers_detector_YOLO3():
filename_image = 'data/ex_detector/bike/Image.png'
path_out = 'data/output/'
if not os.path.exists(path_out):
os.makedirs(path_out)
else:
tools_IO.remove_files(path_out)
CNN = detector_YOLO3.detector_YOLO3(
'./data/ex_yolo/MODELS/model_default.h5',
'./data/ex_YOLO/models/metadata_default.txt')
tools_CNN_view.visualize_layers(CNN.model, filename_image, path_out)
#tools_CNN_view.visualize_filters(CNN.model,path_out)
return
def init_model_16_24(self, input_shape, num_classes):
n_filters = 32 * 3
self.model = Sequential()
self.model.add(BatchNormalization(input_shape=input_shape))
self.model.add(
Conv2D(n_filters, (2, 2),
activation='relu',
strides=(2, 2),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(2, 2)))
N = self.model.layers[-1].output.get_shape().as_list()[3]
self.model.add(
Conv2D(N, (2, 2),
activation='relu',
strides=(2, 2),
padding='same',
data_format='channels_last'))
self.model.add(MaxPooling2D((2, 2), strides=(2, 2)))
self.model.add(Flatten())
self.model.add(Dropout(0.5))
N = numpy.prod(self.model.layers[-3].output.get_shape().as_list()[1:])
self.model.add(Dense(int(N / 4), activation='relu'))
self.model.add(Dense(num_classes, activation='softmax'))
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
tools_CNN_view.visualize_layers(self.model, self.filename_debug_image,
self.folder_debug)
self.model.summary()
return