rtdertv2有可视化pr曲线的代码吗

[xiaocongxin]

Star RTDETR
请先在RTDETR主页点击star以支持本项目
Star RTDETR to help more people discover this project.

---

Describe the bug
A clear and concise description of what the bug is.
If applicable, add screenshots to help explain your problem.

To Reproduce
Steps to reproduce the behavior.
请问rtdertv2有可视化pr曲线的代码吗

[lyuwenyu]

目前没有相关的代码的；

[xiaocongxin]

目前没有相关的代码的；

好的，谢谢。想要绘制混淆矩阵，能否指点一下可以修改哪部分代码

[lyuwenyu]

需要重写一下accumulate；这个函数内部有一个_summarize；可以看到self.eval['precision'], self.eval['recall']怎么使用；不过还是建议使用第三方库直接可以画的那种

RT-DETR/rtdetrv2_pytorch/src/data/dataset/coco_eval.py

Line 73 in 0b6972d

coco_eval.accumulate()

---

https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/cocoeval.py#L411-L418

[philippschw]

Hi, I am struggling with the same issue, did you find a solution already?

[xiaocongxin]

Hi, I am struggling with the same issue, did you find a solution already?

我没有直接画出来，而是将数组保存为npz文件，可能需要你自己修改一下，下面是我修改的部分：
rtdetrv2_pytorch/src/data/dataset/coco_eval.py中的summarize_classwise
` def summarize_classwise(self, coco_eval, iou_type):
precisions = coco_eval.eval['precision']
recalls = coco_eval.eval['recall']
assert len(self.coco_gt.getCatIds()) == precisions.shape[2]

    results_per_category = []
    for idx, cat_id in enumerate(self.coco_gt.getCatIds()):
        np.savez(f'/D/rtdetrv2_pytorch/runs/valnpz/rtdetr_testmetrics_class_{idx}.npz', 
             precisions=precisions[0, :, idx, 0, 2], 
             recalls=np.arange(0.0, 1.01, 0.01))
        t = []
        nm = self.coco_gt.loadCats(cat_id)[0]
        precision = precisions[:, :, idx, 0, -1]
        precision = precision[precision > -1]
        if precision.size:
            ap = np.mean(precision)
        else:
            ap = float('nan')
        t.append(f'{nm["name"]}')
        t.append(f'{round(ap, 3)}')  # 所有阈值下的平均精度

        # indexes of IoU @50 and @50-95

        precision_50 = precisions[0, :, idx, 0, -1]
        precision_50 = precision_50[precision_50 > -1]
        if precision_50.size:
            ap_50 = np.mean(precision_50)
        else:
            ap_50 = float('nan')
        t.append(f'{round(ap_50, 3)}')


        precision_50_95 = precisions[:, :, idx, 0, -1]
        precision_50_95 = precision_50_95[precision_50_95 > -1]
        if precision_50_95.size:
            ap_50_95 = np.mean(precision_50_95)
        else:
            ap_50_95 = float('nan')
        t.append(f'{round(ap_50_95, 3)}')

        # indexes of area of small, median and large
        for area in [1, 2, 3]:
            precision = precisions[:, :, idx, area, -1]
            precision = precision[precision > -1]
            if precision.size:
                ap = np.mean(precision)
            else:
                ap = float('nan')
            t.append(f'{round(ap, 3)}')
        results_per_category.append(tuple(t))

    num_columns = len(results_per_category[0])
    results_flatten = list(itertools.chain(*results_per_category))
    headers = [
        'category', 'mAP', 'mAP_50', 'mAP50-95', 'mAP_s',
        'mAP_m', 'mAP_l'
    ]
    results_2d = itertools.zip_longest(*[
        results_flatten[i::num_columns]
        for i in range(num_columns)
    ])
    table_data = [headers]
    table_data += [result for result in results_2d]
    table = AsciiTable(table_data)
    print(f"\n{iou_type} evaluation per category:\n" + table.table)   `

[xiaocongxin]

Hi, I am struggling with the same issue, did you find a solution already?

要画的话应该就是直接
fig, ax = plt.subplots(1, 1, figsize=(10, 8), tight_layout=True) ax.plot(recalls=np.arange(0.0, 1.01, 0.01), precisions=precisions[0, :, idx, 0, 2], linewidth=3,color="red", label=f"rtdetrv2 {idx} classes {ap_d} [email protected]")

[philippschw]

Thanks for your answer @xiaocongxin .
Because of the different timezones, I already come up with my own solution, maybe it is helpful for someone still. As I did not understand how to call it from inside coco_eval, I load the eval result and from here compute and draw the pr curve and save the [email protected] for each class in a dict for saving it to the model tracker.

def draw_pr_curve(all_precision, save_dir="pr_curve.png"):
    """
    Draws the Precision-Recall curve for each class and calculates [email protected].

    Args:
        all_precision: Array containing precision values for all classes, IoU thresholds, areas, and detection thresholds.
        save_dir: Path to save the plot.

    Returns:
        A dictionary where keys are class names and values are corresponding [email protected] values.

    """
    num_classes = all_precision.shape[2]
    iou_thres_index = 0  # IoU threshold index for [email protected]
    area_range_index = 0  # All Area ranges
    max_detect_thres_index = 2  # Maximal Detection Threshold 100

    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
    colors = plt.colormaps['tab20']

    # Initialize a dictionary to store [email protected] values
    map_05_dict = {}

    for class_id in range(num_classes):
        pr_curve = all_precision[iou_thres_index, :, class_id, area_range_index, max_detect_thres_index]
        x = np.arange(0, 1.01, 0.01)

        # Calculate [email protected]
        map_05 = np.mean(pr_curve)  # Assuming precision is calculated at 101 recall points
        map_05_dict[mscoco_category2name.get(class_id + 1, f'{class_id} not specified')] = map_05

        # Add [email protected] to the legend label
        label = f"{mscoco_category2name.get(class_id + 1, f'{class_id} not specified')} {map_05:.3f}" 
        ax.plot(x, pr_curve, label=label, color=colors(class_id / num_classes))

    # Calculate and store all classes [email protected]
    all_classes_map_05 = np.mean(list(map_05_dict.values()))
    map_05_dict['all_classes'] = all_classes_map_05

    # Add legend
    ax.legend(title='Classes [email protected]', bbox_to_anchor=(1.05, 1), loc='upper left')

    # Add axis descriptions
    ax.set_xlabel('Recall')
    ax.set_ylabel('Precision')
    ax.set_title('Precision-Recall Curve')

    fig.savefig(save_dir, dpi=250)
    plt.close(fig)

    return map_05_dict 

coco_eval = torch.load('output/rtdetrv2_r18vd_120e_coco/eval.pth')
map05 = draw_pr_curve(coco_eval['precision'])

[xiaocongxin]

谢谢你的回答。由于时区不同，我已经想出了自己的解决方案，也许它对某人仍然有帮助。由于我不明白如何从 coco_eval 内部调用它，因此我加载了评估结果，然后从这里计算并绘制 pr 曲线，并将每个类的 [email protected] 保存在一个字典中，以便将其保存到模型跟踪器中。

def draw_pr_curve(all_precision, save_dir="pr_curve.png"):
    """
    Draws the Precision-Recall curve for each class and calculates [email protected].

    Args:
        all_precision: Array containing precision values for all classes, IoU thresholds, areas, and detection thresholds.
        save_dir: Path to save the plot.

    Returns:
        A dictionary where keys are class names and values are corresponding [email protected] values.

    """
    num_classes = all_precision.shape[2]
    iou_thres_index = 0  # IoU threshold index for [email protected]
    area_range_index = 0  # All Area ranges
    max_detect_thres_index = 2  # Maximal Detection Threshold 100

    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
    colors = plt.colormaps['tab20']

    # Initialize a dictionary to store [email protected] values
    map_05_dict = {}

    for class_id in range(num_classes):
        pr_curve = all_precision[iou_thres_index, :, class_id, area_range_index, max_detect_thres_index]
        x = np.arange(0, 1.01, 0.01)

        # Calculate [email protected]
        map_05 = np.mean(pr_curve)  # Assuming precision is calculated at 101 recall points
        map_05_dict[mscoco_category2name.get(class_id + 1, f'{class_id} not specified')] = map_05

        # Add [email protected] to the legend label
        label = f"{mscoco_category2name.get(class_id + 1, f'{class_id} not specified')} {map_05:.3f}" 
        ax.plot(x, pr_curve, label=label, color=colors(class_id / num_classes))

    # Calculate and store all classes [email protected]
    all_classes_map_05 = np.mean(list(map_05_dict.values()))
    map_05_dict['all_classes'] = all_classes_map_05

    # Add legend
    ax.legend(title='Classes [email protected]', bbox_to_anchor=(1.05, 1), loc='upper left')

    # Add axis descriptions
    ax.set_xlabel('Recall')
    ax.set_ylabel('Precision')
    ax.set_title('Precision-Recall Curve')

    fig.savefig(save_dir, dpi=250)
    plt.close(fig)

    return map_05_dict 

coco_eval = torch.load('output/rtdetrv2_r18vd_120e_coco/eval.pth')
map05 = draw_pr_curve(coco_eval['precision'])

好厉害，我是根据
#302
这个问题下Godk02给的代码改的，我不清楚为什么可以这么调用