def SSD512_resnet(input_shape=(512, 512, 3), num_classes=21, softmax=True): # TODO: it does not converge! x = input_tensor = Input(shape=input_shape) source_layers = ssd512_resnet_body(x) # Add multibox head for classification and regression num_priors = [4, 6, 6, 6, 6, 4, 4] normalizations = [20, 20, 20, 20, 20, 20, 20] output_tensor = multibox_head(source_layers, num_priors, num_classes, normalizations, softmax) model = Model(input_tensor, output_tensor) model.num_classes = num_classes # parameters for prior boxes model.image_size = input_shape[:2] model.source_layers = source_layers # stay compatible with caffe models model.aspect_ratios = [[1, 2, 1 / 2], [1, 2, 1 / 2, 3, 1 / 3], [1, 2, 1 / 2, 3, 1 / 3], [1, 2, 1 / 2, 3, 1 / 3], [1, 2, 3, 1 / 2, 1 / 3], [1, 2, 1 / 2], [1, 2, 1 / 2]] model.minmax_sizes = [(35, 76), (76, 153), (153, 230), (230, 307), (307, 384), (384, 460), (460, 537)] model.steps = [8, 16, 32, 64, 128, 256, 512] model.special_ssd_boxes = True return model
def SL512_resnet(input_shape=(512, 512, 3), activation='relu', softmax=True): # body x = input_tensor = Input(shape=input_shape) source_layers = ssd512_resnet_body(x, activation=activation) # Add multibox head for classification and regression num_priors = [1, 1, 1, 1, 1, 1, 1] normalizations = [20, 20, 20, 20, 20, 20, 20] output_tensor = multibox_head(source_layers, num_priors, normalizations, softmax) model = Model(input_tensor, output_tensor) # parameters for prior boxes model.image_size = input_shape[:2] model.source_layers = source_layers return model