mcg-nju / bcn Goto Github PK

Unfortunately I cannot access your pretrained model from the mega link. Can you please share it through google drive or Baidu disk?

Hi, Thanks for sharing the great work.

I am training fullBGM() using your code. I didn't change your code but only added parts to print loss, precision, and recall for both train and val. This is what I got for dataset 50salads.

Basically, val loss keeps increasing and val recall keeps decreasing from the start of training. Is this what should happen? Did I do anything wrong? The increasing loss seems to mean that the model does not learn at all.

It is written in your paper that you select the BGM with the best f1 score. May I ask what that best f1 score is? Based on my training curves so far, it seems that the best f1 score would still be quite a low value.

Thank you.

Hi,

Thanks for sharing the code of your great work.

I noticed that you use n-fold cross validation for all 3 datasets. And I can see from your .bundle files that, in each split, you use 4/5 of all data for training and 1/5 of all data for testing (in the case of 5-fold cv).

I am just wondering how you select the best model, for example, how do you know at which epoch to save the model as your best model to report testing results on the 1/5 test data? Do you further split the 4/5 training data into a train set and a val set and use the small val set for model selection? (Though I do not see this from the code.) Or, do you use the 1/5 test data both for selecting the best model and reporting the results?

Please correct me If I am very wrong in understanding your training process.

Thanks!!

在论文中文字描述和图表中，SC模块中的输入包含置信度分数，但是代码中的描述输入的是所有特征的集合。不知道是不是我分析错误？

您好，你们的工作十分优秀，我想进行复现和进一步学习。
@wanglimin @zhenzhiwang @sebgao
但是把30GB的数据集整体下载比较耗时间，请问能否分别提供三个单独的数据集压缩包？

Hello. Thank you so much for the files and code you uploaded, and we want to see the "Visualization video" in real scene. Can you provide this part of the code?

hi, thx for the great job. But in paper , I can't find batch size config. Was it batch_size=1 which you used in your paper?

Dear authors，

What is the meaning of split? There are many .bundle files in the /data/split, and can you explain this? Thank you!

Hello，I would like to ask how do I determine how many epochs I want to train when pre-training the BGM module.Since it is co-trained with SC again in the later, I don't know how many rounds of pre-training and co-training can achieve better results.Maybe my question is a little too simple, but I still hope to receive your reply, thank you!

Hi,

Have you got code to produce the test loss per epoch for the BCN component?

If not is this better done at training or at prediction point?

Thanks for your contribution!

I used the model you provided for the test. I don't quite understand what these five results represent

Hello, first of all thank you very much for the code and then I would like to ask how can I get this visualization result