Did the classification module help?

[chenyzh28]

Thanks for your work! I removed the classification module and its related loss and the performance is about 77%. I wonder if you have done the experiments with classification loss (it seems to serve as a guide to segmentation, if not detach with previous part).

[chezhizhong]

how do you work it? i work it like you and i get some wrongs: /opt/conda/conda-bld/pytorch_1532502421238/work/aten/src/THCUNN/SpatialClassNLLCriterion.cu:99: void cunn_SpatialClassNLLCriterion_updateOutput_kernel(T *, T *, T *, long *, T *, int, int, int, int, int, long) [with T = float, AccumT = float]: block: [0,0,0], thread: [452,0,0] Assertion t >= 0 && t < n_classes failed. could you give some advice? Thank you!

[chenyzh28]

This problem may be caused by the maximum value in label exceeding n_classes. Have you pre-processed the VOC data following https://www.sun11.me/blog/2018/how-to-use-10582-trainaug-images-on-DeeplabV3-code/

[chezhizhong]

Thanks for your reply! I work my code again, i get a new wrong: pixel_acc += mask_pred.max(dim=1)[1].data.cpu().eq(mask_labels.squeeze(1).cpu()).float().mean(), i check the two's shape, they are identical.

[chenyzh28]

What is the error message?

[chezhizhong]

Traceback (most recent call last):
File "train.py", line 205, in
train(epoch, optimizer, training_loader)
File "train.py", line 131, in train
pixel_acc += mask_pred.max(dim=1)[1].data.cpu().eq(mask_labels.squeeze(1).cpu()).float().mean()
RuntimeError: cuda runtime error (59) : device-side assert triggered at /opt/conda/conda-bld/pytorch_1532502421238/work/aten/src/THC/generic/THCTensorCopy.cpp:70

[chenyzh28]

This is because your GPU has not enough memory for the training process. You can decrease the batch size and try again.

[chezhizhong]

Sorry to bother you again. i get the error again: /opt/conda/conda-bld/pytorch_1532502421238/work/aten/src/THCUNN/SpatialClassNLLCriterion.cu:99: void cunn_SpatialClassNLLCriterion_updateOutput_kernel(T *, T *, T *, long *, T *, int, int, int, int, int, long) [with T = float, AccumT = float]: block: [0,0,0], thread: [228,0,0] Assertion t >= 0 && t < n_classes failed. how can i resolve this error? how the maximum set?

[chenyzh28]

You can follow this link https://www.sun11.me/blog/2018/how-to-use-10582-trainaug-images-on-DeeplabV3-code/ and re-construct your dataset.

[zhenmafan7]

I'm sorry to bother u,I want to ask in dataset.py,
"labels = np.load('/home/liekkas/DISK2/jian/PASCAL/VOC2012/cls_labels.npy')[()]"
I can't find cls_label.npy,what should I do to slove this problem.(明明都是中国人，但是我还是要用渣英文来求助，心好累哦，哭)

[chenyzh28]

这个是分类的标签，作者没有提供，你可以把分类那个分支给去掉，然后这句话可以注释掉，不输出分类的标签。

[zhenmafan7]

这个是分类的标签，作者没有提供，你可以把分类那个分支给去掉，然后这句话可以注释掉，不输出分类的标签。

我知道有点蠢，但是我还是想问，分类的分支是哪个部分啊？TAT

[chenyzh28]

classifier模块（不是mask_classifier）就是分类的分支，你把相关部分都删掉就可以。

[zhenmafan7]

classifier模块（不是mask_classifier）就是分类的分支，你把相关部分都删掉就可以。

谢谢你，我理论上知道了，实际上操作...我慢慢、慢慢、慢慢、慢慢琢磨吧（真难TAT）
thank you
ありがとうございます
감사합니다

[zhenmafan7]

不好意思再次打扰你，我按照你所说把分支模块删除后进行训练，随后运行eval.py时加了print打印了测试结果（如果不加我看不到任何结果，不知道是不是代码哪里改错了），打印的结果如下：
Length of test set:1449 Each_cls_IOU:{'background': 0.0, 'aeroplane': 0.0, 'bicycle': 0.0, 'bird': 0.0, 'boat': 0.0, 'bottle': 0.0, 'bus': 0.0, 'car': 0.0, 'cat': 0.0, 'chair': 0.0, 'cow': 0.0, 'diningtable': 0.0, 'dog': 0.0, 'horse': 0.0, 'motorbike': 0.0, 'person': 18.41086525189786, 'pottedplant': 0.0, 'sheep': 0.0, 'sofa': 0.0, 'train': 0.0, 'tvmonitor': 0.0} mIOU:0.8767 PA:5.03% loss_ic0.000000
这很奇怪，我可以对照一下你修改后的文件吗？

[LeeThrzz]

不好意思再次打扰你，我按照你所说把分支模块删除后进行训练，随后运行eval.py时加了print打印了测试结果（如果不加我看不到任何结果，不知道是不是代码哪里改错了），打印的结果如下：
Length of test set:1449 Each_cls_IOU:{'background': 0.0, 'aeroplane': 0.0, 'bicycle': 0.0, 'bird': 0.0, 'boat': 0.0, 'bottle': 0.0, 'bus': 0.0, 'car': 0.0, 'cat': 0.0, 'chair': 0.0, 'cow': 0.0, 'diningtable': 0.0, 'dog': 0.0, 'horse': 0.0, 'motorbike': 0.0, 'person': 18.41086525189786, 'pottedplant': 0.0, 'sheep': 0.0, 'sofa': 0.0, 'train': 0.0, 'tvmonitor': 0.0} mIOU:0.8767 PA:5.03% loss_ic0.000000
这很奇怪，我可以对照一下你修改后的文件吗？

你好，请问你这个开源代码跑出来了吗，我最近也开始尝试这个代码，遇到了和你一样的问题，在你的个人主页上没有找到邮件，请问可以联系你吗？

[zhenmafan7]

不好意思再次打扰你，我按照你所说把分支模块删除后进行训练，随后运行eval.py时加了print打印了测试结果（如果不加我看不到任何结果，不知道是不是代码哪里改错了），打印的结果如下：
Length of test set:1449 Each_cls_IOU:{'background': 0.0, 'aeroplane': 0.0, 'bicycle': 0.0, 'bird': 0.0, 'boat': 0.0, 'bottle': 0.0, 'bus': 0.0, 'car': 0.0, 'cat': 0.0, 'chair': 0.0, 'cow': 0.0, 'diningtable': 0.0, 'dog': 0.0, 'horse': 0.0, 'motorbike': 0.0, 'person': 18.41086525189786, 'pottedplant': 0.0, 'sheep': 0.0, 'sofa': 0.0, 'train': 0.0, 'tvmonitor': 0.0} mIOU:0.8767 PA:5.03% loss_ic0.000000
这很奇怪，我可以对照一下你修改后的文件吗？

你好，请问你这个开源代码跑出来了吗，我最近也开始尝试这个代码，遇到了和你一样的问题，在你的个人主页上没有找到邮件，请问可以联系你吗？

我训练完之后就是上述的结果，后续没有再尝试过了，主页开放了邮箱可以联系，关于这套代码的话我放弃了，不知道是否还能帮上你。

[FantasyJXF]

if you use torch.nn.CrossEntropyLoss, set the ignore_index may help.

The Pascal VOC ignore classifier 255, which was its white border, if you have boader, just ignore the board ID.

[LiouCZ]

那个分类分支的loss能提高性能吗，有什么用？