eval的结果miou只有1.4好像，请求指导

[ksama777]

您好，我把checkpoint的路径设置成您那个pretrain文件夹里面的checkpoint ，test切片进行， sequence08的eval的结果好像有些不对 您能帮我看看吗 感谢您

这是我test时候的命令 python3 test_SemanticKITTI.py --test_id 08 --result_dir /media/wx/HDD/DQ/RandLA-pytorch-main/result

(randlanet) wx@dl-group-workstation:/media/wx/HDD/DQ/RandLA-pytorch-main$ python3 evaluate_SemanticKITTI.py --dataset /media/wx/HDD/DQ/semantic_kitti/dataset/sequences --predictions /media/wx/HDD/DQ/RandLA-pytorch-main/result --sequence 08

---

INTERFACE:
Eval What: all
Data: /media/wx/HDD/DQ/semantic_kitti/dataset/sequences
Predictions: /media/wx/HDD/DQ/RandLA-pytorch-main/result
Sequences: ['08']
Config: /media/wx/HDD/DQ/RandLA-pytorch-main/utils/semantic-kitti.yaml
Limit: None

---

Opening data config file /media/wx/HDD/DQ/RandLA-pytorch-main/utils/semantic-kitti.yaml
Ignoring xentropy class 0 in IoU evaluation
[IOU EVAL] IGNORE: [0]
[IOU EVAL] INCLUDE: [ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19]
4071
4071 4071
Evaluating sequences
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4071/4071 [02:13<00:00, 30.40it/s]
Validation set:
Acc avg 0.063
IoU avg 0.014
IoU class 1 [car] = 0.017
IoU class 2 [bicycle] = 0.001
IoU class 3 [motorcycle] = 0.000
IoU class 4 [truck] = 0.000
IoU class 5 [other-vehicle] = 0.004
IoU class 6 [person] = 0.001
IoU class 7 [bicyclist] = 0.002
IoU class 8 [motorcyclist] = 0.000
IoU class 9 [road] = 0.090
IoU class 10 [parking] = 0.001
IoU class 11 [sidewalk] = 0.046
IoU class 12 [other-ground] = 0.001
IoU class 13 [building] = 0.009
IoU class 14 [fence] = 0.000
IoU class 15 [vegetation] = 0.012
IoU class 16 [trunk] = 0.000
IoU class 17 [terrain] = 0.079
IoU class 18 [pole] = 0.001
IoU class 19 [traffic-sign] = 0.000

---

0.017,0.001,0.000,0.000,0.004,0.001,0.002,0.000,0.090,0.001,0.046,0.001,0.009,0.000,0.012,0.000,0.079,0.001,0.000,0.014,0.063

[Charles Chen]

请指定test命令的--checkpoint_path 为 pretrain_model/checkpoint.tar 再测试

[ksama777]

您好，用了pretrain的checkpoint&nbsp;结果如下&nbsp; 这么看来是我训练的时候的问题？ (randlanet) ***@***.***:/media/wx/HDD/DQ/RandLA-pytorch-main$ python3 evaluate_SemanticKITTI.py --dataset /media/wx/HDD/DQ/semantic_kitti/dataset/sequences --predictions /media/wx/HDD/DQ/RandLA-pytorch-main/result --sequence 08 ******************************************************************************** INTERFACE: Eval What: all Data:&nbsp; /media/wx/HDD/DQ/semantic_kitti/dataset/sequences Predictions:&nbsp; /media/wx/HDD/DQ/RandLA-pytorch-main/result Sequences:&nbsp; ['08'] Config:&nbsp; /media/wx/HDD/DQ/RandLA-pytorch-main/utils/semantic-kitti.yaml Limit:&nbsp; None ******************************************************************************** Opening data config file /media/wx/HDD/DQ/RandLA-pytorch-main/utils/semantic-kitti.yaml Ignoring xentropy class&nbsp; 0&nbsp; in IoU evaluation [IOU EVAL] IGNORE:&nbsp; [0] [IOU EVAL] INCLUDE:&nbsp; [ 1&nbsp; 2&nbsp; 3&nbsp; 4&nbsp; 5&nbsp; 6&nbsp; 7&nbsp; 8&nbsp; 9 10 11 12 13 14 15 16 17 18 19] 4071 4071 4071 Evaluating sequences 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4071/4071 [01:44<00:00, 38.97it/s] Validation set: Acc avg 0.903 IoU avg 0.550 IoU class 1 [car] = 0.940 IoU class 2 [bicycle] = 0.091 IoU class 3 [motorcycle] = 0.340 IoU class 4 [truck] = 0.666 IoU class 5 [other-vehicle] = 0.460 IoU class 6 [person] = 0.553 IoU class 7 [bicyclist] = 0.687 IoU class 8 [motorcyclist] = 0.000 IoU class 9 [road] = 0.920 IoU class 10 [parking] = 0.401 IoU class 11 [sidewalk] = 0.784 IoU class 12 [other-ground] = 0.005 IoU class 13 [building] = 0.886 IoU class 14 [fence] = 0.519 IoU class 15 [vegetation] = 0.855 IoU class 16 [trunk] = 0.626 IoU class 17 [terrain] = 0.748 IoU class 18 [pole] = 0.567 IoU class 19 [traffic-sign] = 0.399 ******************************************************************************** 0.940,0.091,0.340,0.666,0.460,0.553,0.687,0.000,0.920,0.401,0.784,0.005,0.886,0.519,0.855,0.626,0.748,0.567,0.399,0.550,0.903 ksama ***@***.*** &nbsp;

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "charlesChen02/RandLA-pytorch" ***@***.***&gt;; 发送时间:&nbsp;2022年10月11日(星期二) 下午4:26 ***@***.***&gt;; ***@***.******@***.***&gt;; 主题:&nbsp;Re: [charlesChen02/RandLA-pytorch] eval的结果miou只有1.4好像，请求指导 (Issue #5) 请指定test命令的--checkpoint_path 为 pretrain_model/checkpoint.tar 再测试 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***&gt;

[Charles Chen]

这么看整个模型是没啥问题的, 关于训练, 可以建议几点:

1. batch size尽量调大至最大化利用GPU资源, 可以缩短整体训练时间, 提高效率,
2. epoch 尽量多训练一些, pretrain model是训练的80个左右得到的, 在V100大概用了一整天来训练
3. utils/config.py 中有设定 num_points = 4096 * 10 # Number of input points 代表了一个receptive field 的大小, 一般来说越大效果越好. 如果你在非kitti数据集上训练的话记得调整