PaddleDetection:人脸检测算法细节‘FaceBoxes’

 2021SC@SDUSC

本周分析PaddleDetection特色模型人脸识别里的算法细节FaceBoxes。

初始化函数定义如下

 def __init__(self,
                 backbone="FaceBoxNet",
                 output_decoder=SSDOutputDecoder().__dict__,
                 densities=[[4, 2, 1], [1], [1]],
                 fixed_sizes=[[32., 64., 128.], [256.], [512.]],
                 num_classes=2,
                 steps=[8., 16., 32.]):
        super(FaceBoxes, self).__init__()
        self.backbone = backbone
        self.num_classes = num_classes
        self.output_decoder = output_decoder
        if isinstance(output_decoder, dict):
            self.output_decoder = SSDOutputDecoder(**output_decoder)
        self.densities = densities
        self.fixed_sizes = fixed_sizes
        self.steps = steps

参数说明： backbone：模式训练主干

output_decoder：输出解码器，`SSDOutputDecoder`实例

densities：定义生成的矩阵密度

fixed_sizes：固定大小

num_classes：输出类的数量

step：特征图上相邻优先矩阵的步骤大小

本函数定义了相关输入和输出的格式以及大小，初始化该类的基本参数。

def build(self, feed_vars, mode='train'):
        im = feed_vars['image']
        if mode == 'train':
            gt_bbox = feed_vars['gt_bbox']
            gt_class = feed_vars['gt_class']

        body_feats = self.backbone(im)
        locs, confs, box, box_var = self._multi_box_head(
            inputs=body_feats, image=im, num_classes=self.num_classes)

        if mode == 'train':
            loss = fluid.layers.ssd_loss(
                locs,
                confs,
                gt_bbox,
                gt_class,
                box,
                box_var,
                overlap_threshold=0.35,
                neg_overlap=0.35)
            loss = fluid.layers.reduce_sum(loss)
            return {'loss': loss}
        else:
            pred = self.output_decoder(locs, confs, box, box_var)
            return {'bbox': pred}

build()函数，根据模型定义网络变量，默认为训练模型，并根据参数、输入矩阵等信息调用ssd_loss方法获取损失函数，并将损失函数输出。

def _inputs_def(self, image_shape):
        im_shape = [None] + image_shape
        # yapf: disable
        inputs_def = {
            'image':    {'shape': im_shape,  'dtype': 'float32', 'lod_level': 0},
            'im_id':    {'shape': [None, 1], 'dtype': 'int64',   'lod_level': 0},
            'gt_bbox':  {'shape': [None, 4], 'dtype': 'float32', 'lod_level': 1},
            'gt_class': {'shape': [None, 1], 'dtype': 'int32',   'lod_level': 1},
            'im_shape': {'shape': [None, 3], 'dtype': 'int32',   'lod_level': 0},
        }
        # yapf: enable
        return inputs_def

_input_def()函数定义了数据输入格式，初始化图像大小类型等基础变量。返回初始化完成的数据。

def build_inputs(
            self,
            image_shape=[3, None, None],
            fields=['image', 'im_id', 'gt_bbox', 'gt_class'],  # for train
            use_dataloader=True,
            iterable=False):
        inputs_def = self._inputs_def(image_shape)
        feed_vars = OrderedDict([(key, fluid.data(
            name=key,
            shape=inputs_def[key]['shape'],
            dtype=inputs_def[key]['dtype'],
            lod_level=inputs_def[key]['lod_level'])) for key in fields])
        loader = fluid.io.DataLoader.from_generator(
            feed_list=list(feed_vars.values()),
            capacity=16,
            use_double_buffer=True,
            iterable=iterable) if use_dataloader else None
        return feed_vars, loader

 build_imputs()函数:定义构建模型必要参数：feed_vars、loader并返回

def _multi_box_head(self, inputs, image, num_classes=2):
        def permute_and_reshape(input, last_dim):
            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
            compile_shape = [0, -1, last_dim]
            return fluid.layers.reshape(trans, shape=compile_shape)

        def _is_list_or_tuple_(data):
            return (isinstance(data, list) or isinstance(data, tuple))

        locs, confs = [], []
        boxes, vars = [], []
        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))

        for i, input in enumerate(inputs):
            densities = self.densities[i]
            fixed_sizes = self.fixed_sizes[i]
            box, var = fluid.layers.density_prior_box(
                input,
                image,
                densities=densities,
                fixed_sizes=fixed_sizes,
                fixed_ratios=[1.],
                clip=False,
                offset=0.5,
                steps=[self.steps[i]] * 2)

            num_boxes = box.shape[2]

            box = fluid.layers.reshape(box, shape=[-1, 4])
            var = fluid.layers.reshape(var, shape=[-1, 4])
            num_loc_output = num_boxes * 4
            num_conf_output = num_boxes * num_classes
            # get loc
            mbox_loc = fluid.layers.conv2d(
                input, num_loc_output, 3, 1, 1, bias_attr=b_attr)
            loc = permute_and_reshape(mbox_loc, 4)
            # get conf
            mbox_conf = fluid.layers.conv2d(
                input, num_conf_output, 3, 1, 1, bias_attr=b_attr)
            conf = permute_and_reshape(mbox_conf, 2)

            locs.append(loc)
            confs.append(conf)
            boxes.append(box)
            vars.append(var)

        face_mbox_loc = fluid.layers.concat(locs, axis=1)
        face_mbox_conf = fluid.layers.concat(confs, axis=1)
        prior_boxes = fluid.layers.concat(boxes)
        box_vars = fluid.layers.concat(vars)
        return face_mbox_loc, face_mbox_conf, prior_boxes, box_vars

本函数定义了输入矩阵头配置，并对输入矩阵属性变量进行初始化。