Pointnet++训练记录——Partseg

**参考文章：**https://blog.csdn.net/weixin_44603934/article/details/123589948?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522164844688816782089349638%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=164844688816782089349638&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2_allfirst_rank_ecpm_v1~rank_v31_ecpm-29-123589948.142^v5pc_search_result_control_group,143^v6control&utm_term=pointnet%2B%2B%E5%88%86%E7%B1%BB&spm=1018.2226.3001.4187

数据标注

下载代码，运行案例数据集，没有报错。开始训练自己的数据集。
按照参考文章，进行数据标注。链接：https://blog.csdn.net/weixin_44603934/article/details/123591370?spm=1001.2014.3001.5501
但发现Cloudcompare 导出txt文件时，有的label在第四列，有的在第五列；因此在用作者所给的脚本全部去除第四列数据是有问题的。
由于标注的数据量较少，在进行一个个检查，分为删除第四列、第五列数据后再删除该列。

模型修改：

一开始，我的数据集是 num_class=1，num_part =2，但是运行之后，就会各种报错。
报错1：expected input[16, 135, 2048] to have 130 channels, but got 135 channels instead
修改in_channel = 135后，
报错：

json.decoder.JSONDecodeError: Extra data: line 11 column 1 (char 609)；
raise JSONDecodeError("Expecting value", s, err.value) from None json.decoder.JSONDecodeError: E
block: [0,0,0], thread: [0,0,0] Assertion `t >= 0 && t < n_classes` failed.
1
2
3

实在是崩溃。

还是先按照博主那样， num_class=2，num_part =4
重新修改了数据集，再运行。
报错：

cuDNN error: CUDNN_STATUS_INTERNAL_ERROR
1

百度回答：https://blog.csdn.net/liaoningxinmin/article/details/119139840
回答指向：第一种情况是cuda和pytorch版本不匹配（源代码能成功运行，排除这种情况）；第二种情况模型预测值与真实标签和真实标签不一致。
果然，seg_classes = {‘Circle’: [0, 1], ‘Rectangular’: [2, 3]}
而我在标注的时候，rectangular分类的是[3，4]
修改之后，成功运行！

数据打印

以下是博主提供的打印的部分代码，链接：https://blog.csdn.net/weixin_44603934/article/details/123589948?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522164844688816782089349638%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=164844688816782089349638&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2_allfirst_rank_ecpm_v1~rank_v31_ecpm-29-123589948.142^v5pc_search_result_control_group,143^v6control&utm_term=pointnet%2B%2B%E5%88%86%E7%B1%BB&spm=1018.2226.3001.4187

            for i in range(cur_batch_size):
                aaa = np.argwhere(cur_pred_val[i] == 0)
                for j in aaa:
                    # print(points1[i,j])

                    res1 = open(r'F:\pointnet_pointnet2_pytorch-master\print\05123456_0_' + str(i) + '.txt', 'a')
                    res1.write('\n' + str(points[i, j]).strip('[]'))
                    res1.close()
                    xxxxxx = xxxxxx + 1

                bbb = np.argwhere(cur_pred_val[i] == 1)
                for j in bbb:
                    # print(points1[i, j])

                    res2 = open(r'F:\pointnet_pointnet2_pytorch-master\print\05123456_1_' + str(i) + '.txt', 'a')
                    res2.write('\n' + str(points[i, j]).strip('[]'))
                    res2.close()
                    xxxxxx = xxxxxx + 1

                ccc = np.argwhere(cur_pred_val[i] == 2)
                for j in ccc:
                    # print(points1[i, j])

                    res3 = open(r'F:\pointnet_pointnet2_pytorch-master\print\05654321_2_' + str(i) + '.txt', 'a')
                    res3.write('\n' + str(points[i, j]).strip('[]'))
                    res3.close()
                    xxxxxx = xxxxxx + 1

                ddd = np.argwhere(cur_pred_val[i] == 3)
                for j in ddd:
                    # print(points1[i, j])

                    res4 = open(r'F:\pointnet_pointnet2_pytorch-master\print\05654321_3_' + str(i) + '.txt', 'a')
                    res4.write('\n' + str(points[i, j]).strip('[]'))
                    res4.close()
                    xxxxxx = xxxxxx + 1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

不懂怎么操作，添加到 for i in range(cur_batch_size):后；
报错

 File "test_partseg.py", line 143, in main
    xxxxxx = xxxxxx + 1
UnboundLocalError: local variable 'xxxxxx' referenced before assignment
1
2
3

在前面添加定义xxxxxx =0，继续跑，出错了：

 File "test_partseg.py", line 142, in main
    res3.write('\n' + str(points[i, j]).strip('[]'))
RuntimeError: CUDA error: device-side assert triggered
1
2
3

打印结果

哎，不知道怎么继续