YOLOv8-seg 分割代码详解（三）Val

前言

YOLOv8-seg 分割代码详解（一）Predict
YOLOv8-seg 分割代码详解（二）Train
YOLOv8-seg 分割代码详解（三）Val

  本文主要以源码+注释为主，可以了解 YOLOv8 计算评价指标的具体实现方法。

模型原始输出

preds = model(batch['img'], augment=augment)
preds: (list:2)
	0: (Tensor:(b, 4+cls_n+32, anchors))
	1: (tuple:3)
		0: (list:3)
			0: (Tensor:(b, 64+cls_n, 80, 80))
			1: (Tensor:(b, 64+cls_n, 40, 40))
			2: (Tensor:(b, 64+cls_n, 20, 20))
		1: (Tensor:(b, 32, anchors))
		2: (Tensor:(b, 32, 160, 160))

输出预处理

  NMS，$38=4+\text{class\_score+class}+32$

preds = self.postprocess(preds)
preds: (tuple:2)
	0: (list:b)
		i: (Tensor:(obj_n, 38))
	1: (Tensor:(b, 32, 160, 160))

更新指标

self.update_metrics(preds, batch)


def update_metrics(self, preds, batch):
    """Metrics."""
    """遍历每张图像的输出"""
    for si, (pred, proto) in enumerate(zip(preds[0], preds[1])):
        idx = batch['batch_idx'] == si
        cls = batch['cls'][idx]
        bbox = batch['bboxes'][idx]
        nl, npr = cls.shape[0], pred.shape[0]  # number of labels, predictions
        shape = batch['ori_shape'][si]
        correct_masks = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device)  # init
        correct_bboxes = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device)  # init
        self.seen += 1

        if npr == 0:
            if nl:
                self.stats.append((correct_bboxes, correct_masks, *torch.zeros(
                    (2, 0), device=self.device), cls.squeeze(-1)))
                if self.args.plots:
                    self.confusion_matrix.process_batch(detections=None, labels=cls.squeeze(-1))
            continue

        # Masks
        midx = [si] if self.args.overlap_mask else idx
        gt_masks = batch['masks'][midx]
        pred_masks = self.process(proto, pred[:, 6:], pred[:, :4], shape=batch['img'][si].shape[1:])

        # Predictions
        if self.args.single_cls:
            pred[:, 5] = 0
        predn = pred.clone()
        ops.scale_boxes(batch['img'][si].shape[1:], predn[:, :4], shape,
                        ratio_pad=batch['ratio_pad'][si])  # native-space pred

		"""
		以上为准备工作
		predn: (Tensor:(pn, 38))
		pred_mask: (Tensor:(pn, 160, 160))
		gt_mask: (Tensor:(1, 160, 160))
		"""
        # Evaluate
        if nl:
            height, width = batch['img'].shape[2:]
            tbox = ops.xywh2xyxy(bbox) * torch.tensor(
                (width, height, width, height), device=self.device)  # target boxes
            ops.scale_boxes(batch['img'][si].shape[1:], tbox, shape,
                            ratio_pad=batch['ratio_pad'][si])  # native-space labels
            """
            labelsn: (Tensor:(ln, 5)), 5=cls+xyxy
            correct_bboxes: (Tensor:(pn, 10))
			correct_masks: (Tensor:(pn, 10))
			10=iou0.5~0.95, bool矩阵, 代表在该iou下预测目标是否预测正确
            """
            labelsn = torch.cat((cls, tbox), 1)  # native-space labels
            correct_bboxes = self._process_batch(predn, labelsn)
            correct_masks = self._process_batch(predn,
                                                labelsn,
                                                pred_masks,
                                                gt_masks,
                                                overlap=self.args.overlap_mask,
                                                masks=True)
            if self.args.plots:
                self.confusion_matrix.process_batch(predn, labelsn)

        # Append correct_masks, correct_boxes, pconf, pcls, tcls
        self.stats.append((correct_bboxes, correct_masks, pred[:, 4], pred[:, 5], cls.squeeze(-1)))
        """后续关于绘图和存储的代码省略"""

不同 IoU 阈值下预测框是否准确

def _process_batch(self, detections, labels, pred_masks=None, gt_masks=None, overlap=False, masks=False):
    """
    Return correct prediction matrix
    Arguments:
        detections (array[N, 6]), x1, y1, x2, y2, conf, class
        labels (array[M, 5]), class, x1, y1, x2, y2
    Returns:
        correct (array[N, 10]), for 10 IoU levels
    """
    if masks:
    	"""one-hot"""
        if overlap:
            nl = len(labels)
            index = torch.arange(nl, device=gt_masks.device).view(nl, 1, 1) + 1
            gt_masks = gt_masks.repeat(nl, 1, 1)  # shape(1,640,640) -> (n,640,640)
            gt_masks = torch.where(gt_masks == index, 1.0, 0.0)
        if gt_masks.shape[1:] != pred_masks.shape[1:]:
            gt_masks = F.interpolate(gt_masks[None], pred_masks.shape[1:], mode='bilinear', align_corners=False)[0]
            gt_masks = gt_masks.gt_(0.5)
        iou = mask_iou(gt_masks.view(gt_masks.shape[0], -1), pred_masks.view(pred_masks.shape[0], -1))
    else:  # boxes
    	"""
    	iou: (Tensor:(ln, pn))
    	最普通的iou
    	"""
        iou = box_iou(labels[:, 1:], detections[:, :4])

    return self.match_predictions(detections[:, 5], labels[:, 0], iou)

IoU 细节

  都是最简单的 IoU：交集 / 并集

def box_iou(box1, box2, eps=1e-7):
    """
    Calculate intersection-over-union (IoU) of boxes.
    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
    Based on https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py

    Args:
        box1 (torch.Tensor): A tensor of shape (N, 4) representing N bounding boxes.
        box2 (torch.Tensor): A tensor of shape (M, 4) representing M bounding boxes.
        eps (float, optional): A small value to avoid division by zero. Defaults to 1e-7.

    Returns:
        (torch.Tensor): An NxM tensor containing the pairwise IoU values for every element in box1 and box2.
    """

    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
    (a1, a2), (b1, b2) = box1.unsqueeze(1).chunk(2, 2), box2.unsqueeze(0).chunk(2, 2)
    inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp_(0).prod(2)

    # IoU = inter / (area1 + area2 - inter)
    return inter / ((a2 - a1).prod(2) + (b2 - b1).prod(2) - inter + eps)


def mask_iou(mask1, mask2, eps=1e-7):
    """
    Calculate masks IoU.

    Args:
        mask1 (torch.Tensor): A tensor of shape (N, n) where N is the number of ground truth objects and n is the
                        product of image width and height.
        mask2 (torch.Tensor): A tensor of shape (M, n) where M is the number of predicted objects and n is the
                        product of image width and height.
        eps (float, optional): A small value to avoid division by zero. Defaults to 1e-7.

    Returns:
        (torch.Tensor): A tensor of shape (N, M) representing masks IoU.
    """
    intersection = torch.matmul(mask1, mask2.T).clamp_(0)
    union = (mask1.sum(1)[:, None] + mask2.sum(1)[None]) - intersection  # (area1 + area2) - intersection
    return intersection / (union + eps)

预测框是否准确

def match_predictions(self, pred_classes, true_classes, iou):
    """
    Matches predictions to ground truth objects (pred_classes, true_classes) using IoU.

    Args:
        pred_classes (torch.Tensor): Predicted class indices of shape(N,).
        true_classes (torch.Tensor): Target class indices of shape(M,).
        iou (torch.Tensor): IoU thresholds from 0.50 to 0.95 in space of 0.05.

    Returns:
        (torch.Tensor): Correct tensor of shape(N,10) for 10 IoU thresholds.
    """
    """
    self.iouv: torch.linspace(0.5,0.95,10), 0.5~0.95以0.05间隔
    correct: (Tensor:(pn, 10))
    correct_class: (Tensor:(ln, pn))
    """
    correct = np.zeros((pred_classes.shape[0], self.iouv.shape[0])).astype(bool)
    correct_class = true_classes[:, None] == pred_classes
    for i, iouv in enumerate(self.iouv):
    	"""
    	x: (Tensor:(n, 2))
    	n: 满足 IoU > threshold 且类别匹配的输出个数
    	2: [ln_idx,pn_idx]
    	在这里同一个label会与多个pred匹配, 同一个pred也会和多个label匹配
    	"""
        x = torch.nonzero(iou.ge(iouv) & correct_class)  # IoU > threshold and classes match
        if x.shape[0]:
            # Concatenate [label, detect, iou]
            """matches: (Tensor:(n, 3)), 相当于在 x 每一项后面增添对应的 IoU 数值"""
            matches = torch.cat((x, iou[x[:, 0], x[:, 1]].unsqueeze(1)), 1).cpu().numpy()
            if x.shape[0] > 1:
            	"""根据 IoU 排序以后依次对 pn_idx 和 ln_idx 去重"""
                matches = matches[matches[:, 2].argsort()[::-1]]
                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
            correct[matches[:, 1].astype(int), i] = True
    return torch.tensor(correct, dtype=torch.bool, device=pred_classes.device)

统计指标

  mAP 是对每个类别的 AP 取平均，AP 是 PR 曲线的面积。

def get_stats(self):
    """Returns metrics statistics and results dictionary."""
    """这里把每个batch的结果cat到一起"""
    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*self.stats)]  # to numpy
    if len(stats) and stats[0].any():
        self.metrics.process(*stats)
    """np.bincount: 统计非负整数出现次数, 此处为统计每个类别出现次数"""
    self.nt_per_class = np.bincount(stats[-1].astype(int), minlength=self.nc)  # number of targets per class
    return self.metrics.results_dict

平均精度

def process(self, tp_b, tp_m, conf, pred_cls, target_cls):
    """
    Processes the detection and segmentation metrics over the given set of predictions.

    Args:
        tp_b (list): List of True Positive boxes.
        tp_m (list): List of True Positive masks.
        conf (list): List of confidence scores.
        pred_cls (list): List of predicted classes.
        target_cls (list): List of target classes.
    """
    
    """
    tp_b: (ndarray:(pn,10))
    tp_m: (ndarray:(pn,10))
    conf: (ndarray:(pn))
    pred_cls: (ndarray:(pn))
    target_cls: (ndarray:(ln))
    """
    results_mask = ap_per_class(tp_m,
                                conf,
                                pred_cls,
                                target_cls,
                                plot=self.plot,
                                on_plot=self.on_plot,
                                save_dir=self.save_dir,
                                names=self.names,
                                prefix='Mask')[2:]
    self.seg.nc = len(self.names)
    self.seg.update(results_mask)
    results_box = ap_per_class(tp_b,
                               conf,
                               pred_cls,
                               target_cls,
                               plot=self.plot,
                               on_plot=self.on_plot,
                               save_dir=self.save_dir,
                               names=self.names,
                               prefix='Box')[2:]
    self.box.nc = len(self.names)
    self.box.update(results_box)

def ap_per_class(tp,
                 conf,
                 pred_cls,
                 target_cls,
                 plot=False,
                 on_plot=None,
                 save_dir=Path(),
                 names=(),
                 eps=1e-16,
                 prefix=''):
    """
    Computes the average precision per class for object detection evaluation.

    Args:
        tp (np.ndarray): Binary array indicating whether the detection is correct (True) or not (False).
        conf (np.ndarray): Array of confidence scores of the detections.
        pred_cls (np.ndarray): Array of predicted classes of the detections.
        target_cls (np.ndarray): Array of true classes of the detections.
        plot (bool, optional): Whether to plot PR curves or not. Defaults to False.
        on_plot (func, optional): A callback to pass plots path and data when they are rendered. Defaults to None.
        save_dir (Path, optional): Directory to save the PR curves. Defaults to an empty path.
        names (tuple, optional): Tuple of class names to plot PR curves. Defaults to an empty tuple.
        eps (float, optional): A small value to avoid division by zero. Defaults to 1e-16.
        prefix (str, optional): A prefix string for saving the plot files. Defaults to an empty string.

    Returns:
        (tuple): A tuple of six arrays and one array of unique classes, where:
            tp (np.ndarray): True positive counts for each class.
            fp (np.ndarray): False positive counts for each class.
            p (np.ndarray): Precision values at each confidence threshold.
            r (np.ndarray): Recall values at each confidence threshold.
            f1 (np.ndarray): F1-score values at each confidence threshold.
            ap (np.ndarray): Average precision for each class at different IoU thresholds.
            unique_classes (np.ndarray): An array of unique classes that have data.

    """

    # Sort by objectness
    """按预测分类置信度从大到小排序"""
    i = np.argsort(-conf)
    tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]

    # Find unique classes
    """标签中出现的类别及其对应的目标个数"""
    unique_classes, nt = np.unique(target_cls, return_counts=True)
    nc = unique_classes.shape[0]  # number of classes, number of detections

    # Create Precision-Recall curve and compute AP for each class
    px, py = np.linspace(0, 1, 1000), []  # for plotting
    ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
    for ci, c in enumerate(unique_classes):
        i = pred_cls == c
        n_l = nt[ci]  # number of labels
        n_p = i.sum()  # number of predictions
        if n_p == 0 or n_l == 0:
            continue

        # Accumulate FPs and TPs
        """
        tp: (ndarray:(pn,10)), bool, 代表某个预测框在对应iou下是否正确
        i: (ndarray:(pn,)), bool, 代表预测框类别是否为当前统计类别
        
        fpc: (ndarray:(n,10))
        tpc: (ndarray:(n,10))
        n为当前类别的预测框数量
        cumsum(0)为累加, 每一行是一个预测框, 每一列是一个iou阈值, (i,j)位置的值=j列[0,i]所有数值的和
        由于已经按照conf排序, 这里的结果就是在不同iou下, 不同conf对应的tp和fp数量
        """
        fpc = (1 - tp[i]).cumsum(0)
        tpc = tp[i].cumsum(0)
		
		"""P,R曲线用的是iou=0.5的结果"""
        # Recall
        recall = tpc / (n_l + eps)  # recall curve
        r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0)  # negative x, xp because xp decreases

        # Precision
        precision = tpc / (tpc + fpc)  # precision curve
        p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1)  # p at pr_score

        # AP from recall-precision curve
        for j in range(tp.shape[1]):
            ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
            if plot and j == 0:
                py.append(np.interp(px, mrec, mpre))  # precision at mAP@0.5

    # Compute F1 (harmonic mean of precision and recall)
    f1 = 2 * p * r / (p + r + eps)
    names = [v for k, v in names.items() if k in unique_classes]  # list: only classes that have data
    names = dict(enumerate(names))  # to dict
    if plot:
        plot_pr_curve(px, py, ap, save_dir / f'{prefix}PR_curve.png', names, on_plot=on_plot)
        plot_mc_curve(px, f1, save_dir / f'{prefix}F1_curve.png', names, ylabel='F1', on_plot=on_plot)
        plot_mc_curve(px, p, save_dir / f'{prefix}P_curve.png', names, ylabel='Precision', on_plot=on_plot)
        plot_mc_curve(px, r, save_dir / f'{prefix}R_curve.png', names, ylabel='Recall', on_plot=on_plot)
	
	"""通过f1来选取conf, 决定pr"""	
    i = smooth(f1.mean(0), 0.1).argmax()  # max F1 index
    p, r, f1 = p[:, i], r[:, i], f1[:, i]
    tp = (r * nt).round()  # true positives
    fp = (tp / (p + eps) - tp).round()  # false positives
    return tp, fp, p, r, f1, ap, unique_classes.astype(int)

用 PR 计算 AP

def compute_ap(recall, precision):
    """
    Compute the average precision (AP) given the recall and precision curves.

    Arguments:
        recall (list): The recall curve.
        precision (list): The precision curve.

    Returns:
        (float): Average precision.
        (np.ndarray): Precision envelope curve.
        (np.ndarray): Modified recall curve with sentinel values added at the beginning and end.
    """

    # Append sentinel values to beginning and end
    mrec = np.concatenate(([0.0], recall, [1.0]))
    mpre = np.concatenate(([1.0], precision, [0.0]))

    # Compute the precision envelope
    """
    np.maximum.accumulate(arr)
    返回数组每个元素是包括自身之前的所有元素最大值
    例如: [3, 1, 4, 1, 5, 9, 2, 6, 5] -> [3 3 4 4 5 9 9 9 9]
    precision原本是按conf由高到低获取的, 所以这里要翻转两次
    """
    mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))

    # Integrate area under curve
    method = 'interp'  # methods: 'continuous', 'interp'
    if method == 'interp':
        x = np.linspace(0, 1, 101)  # 101-point interp (COCO)
        """np.trapz: 梯形法则积分"""
        ap = np.trapz(np.interp(x, mrec, mpre), x)  # integrate
    else:  # 'continuous'
        i = np.where(mrec[1:] != mrec[:-1])[0]  # points where x-axis (recall) changes
        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])  # area under curve

    return ap, mpre, mrec

best model 与提前结束训练

  评估当前 epoch 训练好坏的指标是验证集上的 mAP，具体计算如下。训练时可用 patience=0 禁用提前结束训练，默认为50。

fitness = 0.1*(box_map50+seg_map50) + 0.9*(box_map+seg_map)