def create_classifier(architecture,
labels,
input_size,
layers,
dropout,
weights=None,
save_bottleneck=False):
base_model = create_feature_extractor(architecture, input_size, weights)
x = base_model.feature_extractor.outputs[0]
x = GlobalAveragePooling2D()(x)
if len(layers) != 0:
for layer in layers[0:-1]:
x = Dense(layer, activation='relu')(x)
x = Dropout(dropout)(x)
x = Dense(layers[-1], activation='relu')(x)
preds = Dense(len(labels), activation='softmax')(x)
model = Model(inputs=base_model.feature_extractor.inputs[0], outputs=preds)
bottleneck_layer = None
if save_bottleneck:
bottleneck_layer = base_model.feature_extractor.layers[-1].name
network = Classifier(model, input_size, labels, base_model.normalize,
bottleneck_layer)
return network
def create_yolo_network(architecture,
input_size,
nb_classes,
nb_box,
weights):
feature_extractor = create_feature_extractor(architecture, input_size, weights)
yolo_net = YoloNetwork(feature_extractor,
input_size,
nb_classes,
nb_box)
return yolo_net
def representative_dataset_gen(self):
num_imgs = 10
from axelerate.networks.common_utils.feature import create_feature_extractor
backend = create_feature_extractor(self._backend,224)
image_files_list = glob.glob(self._dataset_path + '/**/*.jpg', recursive=True)
for filename in image_files_list[0:10]:
image = cv2.imread(filename)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (self._img_size, self._img_size))
data = np.array(backend.normalize(image),dtype=np.float32)
data = np.expand_dims(data, 0)
yield [data]
def create_classifier(architecture, labels, input_size, layers, dropout):
base_model = create_feature_extractor(architecture, input_size)
x = base_model.feature_extractor.outputs[0]
x = GlobalAveragePooling2D()(x)
for layer in layers[0:-1]:
x = Dense(layer, activation='relu')(x)
x = Dropout(dropout)(x)
x = Dense(layers[-1], activation='relu')(x)
preds = Dense(len(labels), activation='softmax')(x)
model = Model(inputs=base_model.feature_extractor.inputs[0], outputs=preds)
network = Classifier(model, input_size, labels)
return network
def squeezenet_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('SqueezeNet', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
squeezenet[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "vgg_segnet"
return model
def tiny_yolo_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('Tiny Yolo', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
tiny_yolo[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "vgg_segnet"
return model
def resnet50_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('ResNet50', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
resnet50[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "resnet50_segnet"
return model
def full_yolo_segnet(n_classes, input_size=224, encoder_level=3):
encoder = create_feature_extractor('Full Yolo', input_size)
encoder_output = encoder.feature_extractor.layers[
full_yolo[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "vgg_segnet"
return model
def vgg16_segnet(n_classes, input_size=224, encoder_level=3):
encoder = create_feature_extractor('VGG16', input_size)
encoder_output = encoder.feature_extractor.layers[
vgg16[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "vgg_segnet"
return model
def nasnetmobile_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('NASNetMobile', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
nasnetmobile[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "nasnetmobile_segnet"
model.normalize = encoder.normalize
return model
def densenet121_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('DenseNet121', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
densenet121[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "densenet121_segnet"
model.normalize = encoder.normalize
return model
def edgetpu_dataset_gen(self):
num_imgs = None
image_files_list = []
from axelerate.networks.common_utils.feature import create_feature_extractor
backend = create_feature_extractor(self._backend, [self._img_size[0], self._img_size[1]])
image_search = lambda ext : glob.glob(self._dataset_path + ext, recursive=True)
for ext in ['/**/*.jpg', '/**/*.jpeg', '/**/*.png']: image_files_list.extend(image_search(ext))
for filename in image_files_list[:num_imgs]:
image = cv2.imread(filename)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (self._img_size[0], self._img_size[1]))
data = np.array(backend.normalize(image), dtype=np.float32)
data = np.expand_dims(data, 0)
yield [out]
def tiny_yolo_segnet(n_classes, input_size, encoder_level, weights):
encoder = create_feature_extractor('Tiny Yolo', input_size, weights)
encoder_output = encoder.feature_extractor.layers[
tiny_yolo[encoder_level]].output
print(encoder_output)
encoder_input = encoder.feature_extractor.inputs[0]
encoder_level += 1
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "tiny_yolo_segnet"
model.normalize = encoder.normalize
return model
def mobilenet_segnet(n_classes,
input_size=224,
encoder_level=3,
architecture='MobileNet2_5'):
encoder = create_feature_extractor(architecture, input_size)
encoder_output = encoder.feature_extractor.layers[
mobilenet[encoder_level]].output
encoder_input = encoder.feature_extractor.inputs[0]
model = _segnet(n_classes,
encoder_input,
encoder_output,
input_size,
encoder_level=encoder_level)
model.model_name = "mobilenet_segnet"
return model
def k210_dataset_gen(self):
num_imgs = None
image_files_list = []
from axelerate.networks.common_utils.feature import create_feature_extractor
backend = create_feature_extractor(self._backend, [self._img_size[0], self._img_size[1]])
image_search = lambda ext : glob.glob(self._dataset_path + ext, recursive=True)
for ext in ['/**/*.jpg', '/**/*.jpeg', '/**/*.png']: image_files_list.extend(image_search(ext))
temp_folder = os.path.join(os.path.dirname(__file__),'tmp')
os.mkdir(temp_folder)
print(image_files_list)
for filename in image_files_list[:num_imgs]:
image = cv2.imread(filename)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (self._img_size[0], self._img_size[1]))
data = np.array(backend.normalize(image), dtype=np.float32)
data = np.expand_dims(data, 0)
bin_filename = os.path.basename(filename).split('.')[0]+'.bin'
with open(os.path.join(temp_folder, bin_filename), "wb") as f:
data = np.transpose(data, [0, 3, 1, 2])
data.tofile(f)
return temp_folder
