One question about ARSL semi-sup code

[CoderZhangYx]

问题确认 Search before asking

• 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

Hi, thanks for your great work!
When i'm reading your ARSL code, i notice that there exists two config file for 10% seperation training. What does the postfix lsj mean? Which config should i use if i want to reproduce your result?

[nemonameless]

large-scale jittering, you can see the target_sizes of RandomResize are from [400,1333] to [1200, 1333]
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/semi_det/arsl/arsl_fcos_r50_fpn_coco_semi010_lsj.yml#L18
without LSJ, the target_sizes are from [640,1333] to [800, 1333]
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/semi_det/arsl/_base_/arsl_fcos_reader.yml#L5

[CoderZhangYx]

large-scale jittering, you can see the target_sizes of RandomResize are from [400,1333] to [1200, 1333] https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/semi_det/arsl/arsl_fcos_r50_fpn_coco_semi010_lsj.yml#L18 without LSJ, the target_sizes are from [640,1333] to [800, 1333] https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/semi_det/arsl/_base_/arsl_fcos_reader.yml#L5

Thanks for your reply. I was trying to reproduce your result. I ran the code using config file arsl_fcos_r50_fpn_coco_semi010.yml on 8 gpus, but found that mAP stopped gaining at around 0.195 at 80k iters. Any suggestions would be helpful, thanks!

[WWYong3]

你好，请问在ARSL论文中，提到的创新部分JCE和TSA在项目中具体的代码是在哪个位置呢？我找了head 和SSOD损失,都没找到该部分代码，请问您知道吗，非常感谢！

[CoderZhangYx]

你好，请问在ARSL论文中，提到的创新部分JCE和TSA在项目中具体的代码是在哪个位置呢？我找了head 和SSOD损失,都没找到该部分代码，请问您知道吗，非常感谢！

Switch to develop branch, arsl codes are under ppdet/modeling/heads/fcos_head.py and ppdet/modeling/losses/fcos_loss.py.
By the way, tell me your results if you are going to reproduce this work, thanks!