DongSky/lbba_boosted_wsod

train model on other dataset

Opened this issue · 5 comments

fsted commented

Hello, thank you for your paper and code; I want to try to train your model on voc2012 and coco. What should I do.
In addition，i try to run :
bash train_wsod.sh 3 pascal_voc_12 vgg16 voc12_wsddn_pre lbba_final.pth
but found an error :
AssertionError: Number of boxes must match number of ground-truth images
How can I solve it?
Thanks 😄

fsted commented

Since voc2012 does not expose the test set, how can I change the code to train and evaluate the model

DongSky commented

Since voc2012 does not expose the test set, how can I change the code to train and evaluate the model

Hi fsted
For question 2, you can simply comment the line 359 of lbba_boosted_wsod/lib/datasets/pascal_voc.py#L359, this line calls the standard offline evaluation, however, you only need the output files (e.g., VOCdevkit/results/VOC20{07,12)/Main) and submit it to the voc evaluation server.
And for question 1, we did not meet this error during the experiments. We recheck the selective search data and test meta file of VOC 2012, the number of images and box lists are both 10991. To investigate the bug, I need more detailed output.

fsted commented

由于voc2012没有暴露测试集，如何更改代码来训练和评估模型

您好 fsted 对于问题 2，您可以简单地评论 lbba_boosted_wsod/lib/datasets/pascal_voc.py#L359 的第 359 行，此行称为标准离线评估，但是，您只需要输出文件（例如，VOCdevkit/results/VOC20 {07,12)/Main) 并将其提交给 voc 评估服务器。 对于问题 1，我们在实验过程中没有遇到这个错误。我们重新检查了VOC 2012的选择性搜索数据和测试元文件，图像和框列表的数量都是10991。为了调查这个bug，我需要更详细的输出。

Thanks for your reply! I have solved the problem and I managed to reproduce the accuracy in the paper

fsted commented

voc2012 没有公开测试集，如何更改代码来训练和评估模型

您好 fsted 对于问题 2，您可以简单地评论 lbba_boosted_wsod/lib/datasets/pascal_voc.py#L359 的第 359 行，此行称为标准离线评估，但您只需要输出文件（例如，VOCdevkit/results/ VOC20 {07,12)/Main) 并将其给 voc 。 1，在实验过程中我们没有遇到服务器问题。重新检查了 VOC 2012 年的结果我们搜索到的数据和测试元文件，对于图像的结果和框列表的数量都是10991。为了更查看这个bug，我需要详细的输出。

by the way,I would like to ask, does the project contain training code for the adjuster?

DongSky commented

voc2012 没有公开测试集，如何更改代码来训练和评估模型

您好 fsted 对于问题 2，您可以简单地评论 lbba_boosted_wsod/lib/datasets/pascal_voc.py#L359 的第 359 行，此行称为标准离线评估，但您只需要输出文件（例如，VOCdevkit/results/ VOC20 {07,12)/Main) 并将其给 voc 。 1，在实验过程中我们没有遇到服务器问题。重新检查了 VOC 2012 年的结果我们搜索到的数据和测试元文件，对于图像的结果和框列表的数量都是10991。为了更查看这个bug，我需要详细的输出。

by the way,I would like to ask, does the project contain training code for the adjuster?

We have released the template code of initialization step, E-step and M-step (to train the adjuster) in the google drive.
Actually, this code is based on https://github.com/ruotianluo/pytorch-faster-rcnn , so you can also build your own adjuster training code maunally.
See https://drive.google.com/drive/folders/1rlJlT34l5mK7h0tmb-bsnPqPM_YqdMgR?usp=sharing for more details.