Open Be997398715 opened 2 years ago
Hi. Could provide more details in the training? For example, the curve of each term of loss in Tensorboard.
Here are some simple suggestions for debugging your code:
1, Try to replace the self-supervision loss with supervision loss in the train.py
, and use train.sh
to train the model. If this works fine, then the problem must locate in the self-supervision loss.
2, Try to use only the photometric consistency loss in the train.py
, and use train.sh
to train the model. If this works, then the problem must locate in the other losses in JDACS.
3, The rest can be done in the same manner, until you find what code leads to the corruption in self-supervision training.
4, My empirical experience is that the augmentation loss sometimes leads to corrupt in self-supervision training, because when the data-augmentation is too hard, the self-supervision may fall into trivial solution.
Hi, I try to use only the photometric consistency loss in the train.py(comment the co-seg loss backward and augment loss backward), and use only one scale for training, these are what summary looks like:
And the images are like:
Is there any problems? Looking forward to your reply. ------------------ 原始邮件 ------------------ 发件人: "ToughStoneX/Self-Supervised-MVS" @.>; 发送时间: 2022年2月13日(星期天) 晚上8:15 @.>; @.**@.>; 主题: Re: [ToughStoneX/Self-Supervised-MVS] JDACS-MS for replacing cvp-mvsnet to patchmatchnet (Issue #12)
Hi. Could provide more details in the training? For example, the curve of each term of loss in Tensorboard. Here are some simple suggestions for debugging your code: 1, Try to replace the self-supervision loss with supervision loss in the train.py, and use train.sh to train the model. If this works fine, then the problem must locate in the self-supervision loss. 2, Try to use only the photometric consistency loss in the train.py, and use train.sh to train the model. If this works, then the problem must locate in the other losses in JDACS. 3, The rest can be done in the same manner, until you find what code leads to the corruption in self-supervision training. 4, My empirical experience is that the augmentation loss sometimes leads to corrupt in self-supervision training, because when the data-augmentation is too hard, the self-supervision may fall into trivial solution.
— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>
Sorry, I cannot see the images on the webpage. Maybe you can send these images to my e-mail? Here is my e-mail: HoBinX@163.com.
I have send these pictures, please check when you are free.
------------------ 原始邮件 ------------------ 发件人: "ToughStoneX/Self-Supervised-MVS" @.>; 发送时间: 2022年2月16日(星期三) 下午3:29 @.>; @.**@.>; 主题: Re: [ToughStoneX/Self-Supervised-MVS] JDACS-MS for replacing cvp-mvsnet to patchmatchnet (Issue #12)
Sorry, I cannot see the images on the webpage. Maybe you can send these images to my e-mail? Here is my e-mail: @.***
— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>
I have send these pictures, please check when you are free. … ------------------ 原始邮件 ------------------ 发件人: "ToughStoneX/Self-Supervised-MVS" @.>; 发送时间: 2022年2月16日(星期三) 下午3:29 @.>; @.**@.>; 主题: Re: [ToughStoneX/Self-Supervised-MVS] JDACS-MS for replacing cvp-mvsnet to patchmatchnet (Issue #12) Sorry, I cannot see the images on the webpage. Maybe you can send these images to my e-mail? Here is my e-mail: @. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.>
Hi, im trying to do the same replacement, have you found your modification durable?
Hi, thanks for your great job! I'd like to replace cvp-mvsnet to patchmatchnet and train the new model. However I found the model loss can't converge to a low level(about 30), is there any suggestions for train a new model? The training strategies are same as train.sh your provided(the only one change is scale set from 2 to 3).