RuntimeError: The size of tensor a (96774) must match the size of tensor b (290322) at non-singleton dimension 0

[lamhoangtung]

I'm trying to rerun your code but I encounter this:

linus@srv-aws:~/2DOCR/ultra_high_resolution_segmentation$ ./train_deep_globe.sh
fpn_global.508_4.28.2019_lr2e5
mode: 1 evaluation: False test: False
preparing datasets and dataloaders......
creating models......
Using poly LR Scheduler!
start training......
  0%|                                                                                                                                                                                                                 | 0/215 [00:00<?, ?it/s]
=>Epoches 0, learning rate = 0.0000500,                 previous best = 0.0000
Traceback (most recent call last):
  File "train_deep_globe.py", line 105, in <module>
    loss = trainer.train(sample_batched, model, global_fixed)
  File "/mnt/data/linus/2DOCR/ultra_high_resolution_segmentation/helper.py", line 346, in train
    loss = self.criterion(outputs_global, labels_glb)
  File "train_deep_globe.py", line 85, in <lambda>
    criterion = lambda x,y: criterion1(x, y)
  File "/mnt/data/linus/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/mnt/data/linus/2DOCR/ultra_high_resolution_segmentation/utils/loss.py", line 57, in forward
    probs = (probs * target).sum(1)
RuntimeError: The size of tensor a (96774) must match the size of tensor b (290322) at non-singleton dimension 0

Any ideal how to fix this @chenwydj ? Thanks a lot.

[lamhoangtung]

Hi @chenwydj, thanks for your amazing works. I'm trying to rerun your experiments on DeepGlobe.
Since I can't find any mask label for the validation set and test set, I splited some sample from training set for validation and testing.
Currently I'm facing the issue above: RuntimeError: The size of tensor a (96774) must match the size of tensor b (290322) at non-singleton dimension 0.
After debugging for a while I found that the input and the target for focal loss calculation don't matching shape, the predicted tensor have 7 channel but the label mask only have 3.
For now, I'm trying to fix it by writing my own code to create the labels mask for loss calculation. You can see my code in commit 0cad452. I really surprised that you have the code to do that but haven't use it anywhere ??? So I did leverage it a lot ;)
But up to this point I'm facing a new issue like this:

Traceback (most recent call last):
  File "train_deep_globe.py", line 129, in <module>
    loss = trainer.train(sample_batched, model, global_fixed)
  File "/root/ultra_high_resolution_segmentation/helper.py", line 429, in train
    self.metrics_global.update(labels_npy, predictions_global)
  File "/root/ultra_high_resolution_segmentation/utils/metrics.py", line 23, in update
    tmp = self._fast_hist(lt.flatten(), lp.flatten(), self.n_classes)
  File "/root/ultra_high_resolution_segmentation/utils/metrics.py", line 18, in _fast_hist
    hist = np.bincount(n_class * label_true[mask].astype(int) + label_pred[mask], minlength=n_class**2).reshape(n_class, n_class)
IndexError: boolean index did not match indexed array along dimension 0; dimension is 5992704 but corresponding boolean dimension is 17978112

So I've been thinking my fix might not be how thing should be ...
Do you have any ideal to resolve this ... Can you take a look at the code (it's all available here: https://github.com/lamhoangtung/ultra_high_resolution_segmentation/tree/deepglobe, on deepglobe branch)

Thanks ;)

[zxshi]

I have encountered the same problem, have you solved it?

[lamhoangtung]

@zxshi I haven't :3 ...
And also no respond from the author :(( ...

[bigKoki]

have you solved it?

[lamhoangtung]

have you solved it?

Solved for binary segmantation

[bigKoki]

can you tell me how to solve it？do you still use this code？

[lamhoangtung]

can you tell me how to solve it？do you still use this code？

I solve it by making sure that the predicted tensor tensor and the target mask tensor have the same number of channel.

You can see the code at https://github.com/lamhoangtung/ultra_high_resolution_segmentation/tree/master

But keep in mind this only work for binary semantic segmantation yet

[bigKoki]

can you tell me how to solve it？do you still use this code？

I solve it by making sure that the predicted tensor tensor and the target mask tensor have the same number of channel.

You can see the code at https://github.com/lamhoangtung/ultra_high_resolution_segmentation/tree/master

But keep in mind this only work for binary semantic segmantation yet

thank you very much！！！it will help

[zzx0836]

@lamhoangtung when I train mode=3 I have a problem:AttributeError: 'Trainer' object has no attribute 'template'?
Can you give me some help?
thanks

[bigKoki]

hi,I want to study this code,and i didn't get a profound comprehension to the Semantic segmentation ,could you please explain the function and distinguish in mode(1,2,3),thank you very much!

[bigKoki]

Hi GLnet is a segmented network model that combine global information with local information. It is divided into global and local networks. In the process of training, interaction between global information and local information fusion. mode=1:train global mode=2:train local form global mode=3 train global from local.

你好，我想你应该是中国人吧，我最近才看这个，知道这个网络结构主要是由全局信息融合局部信息，但是在代码里看到了有三种mode，这三种mode是网络结构不同吗，还是训练的结果不一样？
麻烦您能讲解一下吗，多谢了！

[bigKoki]

@lamhoangtung when I train mode=3 I have a problem:AttributeError: 'Trainer' object has no attribute 'template'?
Can you give me some help?
thanks

多谢讲解！！！让我有了进一步的理解

[lamhoangtung]

@lamhoangtung when I train mode=3 I have a problem:AttributeError: 'Trainer' object has no attribute 'template'?
Can you give me some help?
thanks

I'm trying to train with mode 3 now but I'm getting NAN loss because the forward pass are giving me NAN tensor.

Still looking into it for now ;). Code at the master branch of my fork

[lamhoangtung]

OK,Thanks.If you solve all the problems, please provide the details of the training, thank you very much ------------------ 原始邮件 ------------------ 发件人: "Hoàng Tùng Lâm (Linus)"notifications@github.com; 发送时间: 2019年8月2日(星期五) 晚上6:23 收件人: "chenwydj/ultra_high_resolution_segmentation"ultra_high_resolution_segmentation@noreply.github.com; 抄送: "1416055750"1416055750@qq.com;"Comment"comment@noreply.github.com; 主题: Re: [chenwydj/ultra_high_resolution_segmentation] RuntimeError: Thesize of tensor a (96774) must match the size of tensor b (290322) atnon-singleton dimension 0 (#6) @lamhoangtung when I train mode=3 I have a problem:AttributeError: 'Trainer' object has no attribute 'template'? Can you give me some help? thanks I'm trying to train with mode 3 now but I'm getting NAN loss because the forward pass are giving me NAN tensor. Still looking into it for now ;). Code at the master branch of my fork — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

Use my fork code, branch master ;)

[lamhoangtung]

@zzx0836 I fixed that in e3c5128 ;). Mode 3 are going well now :P

[zzx0836]

thank you very much !

…

------------------ 原始邮件 ------------------ 发件人: "Hoàng Tùng Lâm (Linus)"<notifications@github.com>; 发送时间: 2019年8月3日(星期六) 凌晨1:15 收件人: "chenwydj/ultra_high_resolution_segmentation"<ultra_high_resolution_segmentation@noreply.github.com>; 抄送: "1416055750"<1416055750@qq.com>;"Mention"<mention@noreply.github.com>; 主题: Re: [chenwydj/ultra_high_resolution_segmentation] RuntimeError: Thesize of tensor a (96774) must match the size of tensor b (290322) atnon-singleton dimension 0 (#6) @zzx0836 I fixed that in e3c5128 ;). Mode 3 are going well now :P — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[wangbyz]

Hi, have you fix this problem ?
I want to run this code in mode 1 but meet the same problem.

[bigKoki]

hi,could you please tell me which dataset you used in this code,i see that the author didnt provide the training code for AerialImageDataset。

[zzx0836]

I'm using my own dataset, and the data is fundus data.

…

------------------ 原始邮件 ------------------ 发件人: "湖北工业停水停电大学"<notifications@github.com>; 发送时间: 2019年8月23日(星期五) 上午10:30 收件人: "TAMU-VITA/GLNet"<GLNet@noreply.github.com>; 抄送: "1416055750"<1416055750@qq.com>;"Mention"<mention@noreply.github.com>; 主题: Re: [TAMU-VITA/GLNet] RuntimeError: The size of tensor a (96774) mustmatch the size of tensor b (290322) at non-singleton dimension 0 (#6) hi,could you please tell me which dataset you used in this code,i see that the author didnt provide the training code for AerialImageDataset。 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[bigKoki]

@zzx0836，thank you，i use the code which forked by lumhoangtung，and it woks

[zzx0836]

You're welcome!

…

------------------ 原始邮件 ------------------ 发件人: "湖北工业停水停电大学"<notifications@github.com>; 发送时间: 2019年8月30日(星期五) 下午3:00 收件人: "TAMU-VITA/GLNet"<GLNet@noreply.github.com>; 抄送: "1416055750"<1416055750@qq.com>;"Mention"<mention@noreply.github.com>; 主题: Re: [TAMU-VITA/GLNet] RuntimeError: The size of tensor a (96774) mustmatch the size of tensor b (290322) at non-singleton dimension 0 (#6) @zzx0836，thank you，i use the code which forked by lumhoangtung，and it woks — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[bigKoki]

hi,can you run it on deepglobe dataset? I have used your code on AerialImageDataset,but failed on deepglobe

[lamhoangtung]

hi,can you run it on deepglobe dataset? I have used your code on AerialImageDataset,but failed on deepglobe

No I can't :3

[chenwydj]

Hi! We updated the instructions for training and evaluation in the Readme, and also uploaded the pretrained model for the Deep Globe. The updated training and evaluation bash scripts work on my side. Please take a look and give a try! :)

[lamhoangtung]

Thanks a lot @chenwydj. It worked ;)