mxnet复现SSD之损失、评价函数

mxnet复现SSD系列文章目录

一、数据集的导入.
二、SSD模型架构.
三、训练脚本的实现.
四、损失、评价函数.
五、预测结果.

---

---

前言

本项目是按照pascal voc的格式读取数据集，数据集为kaggle官网提供的口罩检测数据集，地址：Face Mask Detection，模型架构参考自gluoncv ssd_300_vgg16_atrous_voc源码

---

一、损失函数

类别损失函数采用SoftmaxCrossEntropyLoss
boundingbox损失函数采用smooth_l1或focal loss

代码实现

from mxnet.gluon.loss import Loss


class smooth_l1(Loss):
    def __init__(self, weight=None, batch_axis=0, **kwargs):
        super(smooth_l1, self).__init__(weight, batch_axis, **kwargs)

    def hybrid_forward(self, F, pred, label):
        loss = F.smooth_l1(pred-label, scalar=1.)
        return F.mean(loss, axis=self._batch_axis, exclude=True)


class FocalLoss(Loss):
    def __init__(self,axis=-1,alpha=0.25,gamma=2,batch_axis=0,**kwargs):
        super(FocalLoss,self).__init__(None,batch_axis,**kwargs)
        self.alpha = alpha
        self.gamma = gamma
        self.axis = axis
        self.batch_axis = batch_axis

    def hybrid_forward(self, F, y, label):
        y = F.softmax(y)
        pt = F.pick(y, label, axis=self.axis, keepdims=True)
        loss = -self.alpha * ((1 - pt) ** self.gamma) * F.log(pt)
        return F.mean(loss, axis=self._batch_axis, exclude=True)

二、评价函数

评价函数采用计算每个类别的recall, precision和AP值

代码实现

class_recs = {}
    npos = 0
    for imagename in imagenames:
        R = [obj for obj in recs[imagename] if obj['name'] == classname] 
                bbox = np.array([x['bbox'] for x in R])
        difficult = np.array([x['difficult'] for x in R]).astype(np.bool)
        det = [False] * len(R) #这个值是用来判断是否重复检测的
        npos = npos + sum(~difficult)
        class_recs[imagename] = {'bbox': bbox,
                                 'difficult': difficult,
                                 'det': det}

    # read dets
    detfile = detpath.format(classname)
    with open(detfile, 'r') as f:
        lines = f.readlines()

    splitlines = [x.strip().split(' ') for x in lines]
    image_ids = [x[0] for x in splitlines]
    confidence = np.array([float(x[1]) for x in splitlines])
    BB = np.array([[float(z) for z in x[2:]] for x in splitlines])

    # sort by confidence 
    sorted_ind = np.argsort(-confidence)
    BB = BB[sorted_ind, :]
    image_ids = [image_ids[x] for x in sorted_ind]

    # go down dets and mark TPs and FPs
    nd = len(image_ids) 
    tp = np.zeros(nd)
    fp = np.zeros(nd)
    for d in range(nd):
        R = class_recs[image_ids[d]]
        bb = BB[d, :].astype(float)
        ovmax = -np.inf 
        BBGT = R['bbox'].astype(float)

        if BBGT.size > 0:
            # compute overlaps
            # intersection
            ixmin = np.maximum(BBGT[:, 0], bb[0])
            iymin = np.maximum(BBGT[:, 1], bb[1])
            ixmax = np.minimum(BBGT[:, 2], bb[2])
            iymax = np.minimum(BBGT[:, 3], bb[3])
            iw = np.maximum(ixmax - ixmin + 1., 0.)
            ih = np.maximum(iymax - iymin + 1., 0.)
            inters = iw * ih

            # union
            uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
                   (BBGT[:, 2] - BBGT[:, 0] + 1.) *
                   (BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)

            overlaps = inters / uni
            ovmax = np.max(overlaps)
            jmax = np.argmax(overlaps)

        if ovmax > ovthresh:
            if not R['difficult'][jmax]: 
                if not R['det'][jmax]:
                    tp[d] = 1.
                    R['det'][jmax] = 1 #判断是否重复检测，检测过一次以后，值就从False变为1了
                else:
                    fp[d] = 1.
        else:
            fp[d] = 1.

    # compute precision recall
    fp = np.cumsum(fp) 
    tp = np.cumsum(tp)
    rec = tp / float(npos)
    # avoid divide by zero in case the first detection matches a difficult
    # ground truth
    prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
    ap = voc_ap(rec, prec, use_07_metric)

    return rec, prec, ap

计算AP值

def voc_ap(rec, prec, use_07_metric=False):
    """Compute VOC AP given precision and recall. If use_07_metric is true, uses
    the VOC 07 11-point method (default:False).
    """
    if use_07_metric:
        # 11 point metric
        ap = 0.
        for t in np.arange(0., 1.1, 0.1):
            if np.sum(rec >= t) == 0:
                p = 0 
            else:
                p = np.max(prec[rec >= t])
            ap = ap + p / 11.
    else:
        # correct AP calculation
        # first append sentinel values at the end
        mrec = np.concatenate(([0.], rec, [1.]))
        mpre = np.concatenate(([0.], prec, [0.]))

        # compute the precision envelope
        for i in range(mpre.size - 1, 0, -1):
            mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
        
        i = np.where(mrec[1:] != mrec[:-1])[0]

        # and sum (\Delta recall) * prec
        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) #计算面积
    return ap

计算mAP值

def mAP():

    detpath,annopath,imagesetfile,cachedir,class_path = get_dir('kitti')
    ovthresh=0.3,
    use_07_metric=False

    rec = 0; prec = 0; mAP = 0
    class_list = get_classlist(class_path)
    for classname in class_list:
        rec, prec, ap = voc_eval(detpath,
                                 annopath,
                                 imagesetfile,
                                 classname,
                                 cachedir,
                                 ovthresh=0.5,
                                 use_07_metric=False,
                                 kitti=True)
        print('on {}, the ap is {}, recall is {}, precision is {}'.format(classname, ap, rec[-1], prec[-1]))
        mAP += ap
    
    mAP = float(mAP) / len(class_list)

    return mAP

参考链接

https://zhuanlan.zhihu.com/p/43068926