ggsding / bi-srnet Goto Github PK

An error is reported when running, but there is no other output, how can I solve it?

Hello, your method is for semantic change detection. Can we use binary change datasets for binary change detection? Hope to get your answer, thank you and wish you a happy life!

Hello, I use the SECOND dataset like you. How do you divide the number of training datasets and validation datasets? There are a total of 2968 images, of which 2500 are used as training data and 468 as verification data sets, how do you deal with them?

Can you share the code based on the framework of DSCD-e, DSCD-l and SSCD-e?can you share Implementation code of these methods，such FC-EF...and HRSCD?
Thank you very much!

I have tried to use the default hyperparameters to train the posted model on SECOND dataset and I'm sure my dataset preparation is fine. Unluckily, it seems that the model fails to learn well and the binary change map generated by the network is far from satisfaction.
(the listed images are binary-change-map, change-detection-result-t1 and change-detection-result-t2)



Maybe there are some settings in the scripts I failed to modify.
Could you offer me a pretrained.pth or some fine-tuned .pth checkpoint as in train_SCD.py line 55 that could help me out?
Many thanks.

what does Fscd which referred in Table 1、 Table 2 and Table 3 mean?

Is it the Score ?
At the time of my reproduction, I didn't get the evaluation index of the name.

Hi, in the BiSRNet.py, lines 149 and 154 do not correspond to the formulas (16)-(19) in the paper. Should it be changed to "energy1 = torch.bmm (q1, k1)" and "energy2 = torch.bmm (q2, k2)"?

Hi, @ggsDing . I learned a lot from your work. I want to know whether you can share the implementation of the methods you compared (your modified FC-EF, FC-Siam-conc, HRSCD, IFN), Thank you very much.

Hi, I just got a bunch of totally blank images in reproducing the process , could you please tell me why?

Loading the dataset takes up too much memory, resulting in frequent training stops, 512 × The image size is 512, and the batch size is 8.

Excuse me， I'm new and would like to get your prediction script to reproduce the full content. I have tried to complete the prediction script by myself but failed.

hello, could you make your pre trained_model public?

在将数据集划分为训练集以及测试集的时候，验证集占比多少，是按数据集的顺序划分的还是随机抽取划分的

您好，在使用SSCDl作为Net训练时，没有问题。在使用BiSRNet作为Net训练时，出现了问题，参数和数据及数据读取方式全部是按照您的代码做的。
出现的问题是：
训练的时候显示：WARNING:Nan or Inf found in Input tensor，且train_seg loss bn_loss均为nan

您好！我用的您发的源码，默认的超参数，但是跑出来的效果比较差，您能提供一下您调整过后的超参数吗？谢谢啦

您好，我尝试用您的代码训练的时候出现了如下问题。似乎是在计算loss的时候target应该是三维，但是debug进去发现label却还是四维，不知道该如何解决呢

Traceback (most recent call last):
File "/data/yuanyifeng/Bi-SRNet-main/train_SCD.py", line 238, in
main()
File "/data/yuanyifeng/Bi-SRNet-main/train_SCD.py", line 65, in main
train(train_loader, net, criterion, optimizer, scheduler, args, val_loader)
File "/data/yuanyifeng/Bi-SRNet-main/train_SCD.py", line 105, in train
loss_seg = criterion(outputs_A, labels_A) * 0.5 + criterion(outputs_B, labels_B) * 0.5
File "/home/yuanyifeng/anaconda3/envs/ASN3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/data/yuanyifeng/Bi-SRNet-main/utils/loss.py", line 13, in forward
return self.nll_loss(F.log_softmax(inputs, dim=1), targets)
File "/home/yuanyifeng/anaconda3/envs/ASN3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/yuanyifeng/anaconda3/envs/ASN3/lib/python3.7/site-packages/torch/nn/modules/loss.py", line 213, in forward
return F.nll_loss(input, target, weight=self.weight, ignore_index=self.ignore_index, reduction=self.reduction)
File "/home/yuanyifeng/anaconda3/envs/ASN3/lib/python3.7/site-packages/torch/nn/functional.py", line 2266, in nll_loss
ret = torch._C._nn.nll_loss2d(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
RuntimeError: 1only batches of spatial targets supported (3D tensors) but got targets of size: : [2, 512, 512, 3]

Hello, this is a very meaningful work, but I have a question: may I ask you how to set the random seed? I tried to add different random seed functions before the training code, but the results of each training still changed, is it convenient for you to share the random seed functions in this project?

How should I modify the code to resolve this error？

您好，SSCDL使用的是SGD和0.1的初始学习率，这个学习率即使是对于SGD来说也是一个比较大的初始值，但在您的网络上有着比较良好的性能，如果修改网络这个学习率将没办法发挥网络的全部性能，（即使是是使用Bisrnet时我也发现了这个问题），想请问您对这个问题有什么见解吗？还是说SGD+0.1仅仅是与SSCDL最匹配，在使用Bisrnet或者别的网络的时候还需要更换学习率和策略？

Hello! 我在运行Eval_SCD.py时，使用second数据集，输入pred_SCD生成的Inference RGB image（im1_rgb），遇到错误
assert unique_set.issubset(set([0, 1, 2, 3, 4, 5, 6])), "unrecognized label number" AssertionError: unrecognized label number，
请问EVAL（）函数中INFER_DIR1和LABEL_DIR1的输入应该是什么？Thanks！

Hello, I read this issue and I'm not going well with it, too:
#21

Can you explain in more detail what parts of the code I should change?

Hello, I tried to test your proposed model on the Landsat-SCD dataset, but it didn't work well, I tried making the following changes:
num_classes = 5
ST_COLORMAP = [[255,255,255], [128,128,128], [130,87,87], [255,0,0], [0,0,255]]
ST_CLASSES = ['unchanged', 'farmland', 'desert', 'building', 'water']
The above modifications occur in : ./ datasets/RS_ST.py.
I don't know if you have done experiments on this dataset, can you give me a little comment?

你好！运行train_SCD.py时加载完数据后执行train函数报MemoryError，请问是什么问题？

The normal data set should include the training set verification set and the test set, and the final performance indicators are calculated on the test set, but the test set of the SECOND data set has no label, only two time-phase images.
Are the indicators for measuring network performance obtained from the verification set?
Or your complete data, that is, the label of the test set

The images contained in these two folders are the train and val in my folder. Do you mean to use the data in val.txt to test and output evaluation indicators?
If you need to divide the images contained in train.txt into training images and validation images, what are train and val images?

i notice that in the paper you used second database to train network, but i did't found it, could you share it or give me a website?

Hi,
I have used your original code and run the pred_SCD.py,however, the result only has four folders:im1, im1_rgb, im2 and im2_rgb.Why has no change map in my results?Could you please give me some specific help? Thank you in advance!