Closed nevynwong closed 1 year ago
Hi, we use some data augmentation methods (e.g., random flip,random rotation and random crop) to avoid overfitting problem. Besides, the weight_decay of the AdamW optimizer is set at 0.7 to avoid overfitting.
Yes, thanks for your prompt reply, I finished the experiment using your publicly available code and found the problem, maybe it's my processing problem.
Hello, may I ask you how to run this author's code? I am a newcomer, his README is too rough, so I would like to ask you how to run this code. This is my email address:2580600151@qq.com
Hi, we use some data augmentation methods (e.g., random flip,random rotation and random crop) to avoid overfitting problem. Besides, the weight_decay of the AdamW optimizer is set at 0.7 to avoid overfitting.
I have another question, the article says to avoid overfitting, randomly select four frames around the onset frame and offset frame, but I found that these frames are not used in the network, I wonder how these are used to avoid overfitting? That means a lot to me. Thank you
Hi, ''randomly select four frames around the onset frame and offset frame'' means we select the onset frame from the four frames around it, we only use one of the four frames at a time.
在2022-10-17 @.***写道:
Hi, we use some data augmentation methods (e.g., random flip,random rotation and random crop) to avoid overfitting problem. Besides, the weight_decay of the AdamW optimizer is set at 0.7 to avoid overfitting.
I have another question, the article says to avoid overfitting, randomly select four frames around the onset frame and offset frame, but I found that these frames are not used in the network, I wonder how these are used to avoid overfitting? That means a lot to me. Thank you
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>
This work may suffer from a data leakage issues as in https://github.com/Justin900429/Facial-Separated-Micro-Expression-Recognition.
Haven't run the code but the lines 580 and 581 seem to indicate that https://github.com/muse1998/MMNet/blob/623e4b81d5de26d5354ffac1757c5af4afcc8d15/CASME2_3.py#L580-L581
This work may suffer from a data leakage issues as in https://github.com/Justin900429/Facial-Separated-Micro-Expression-Recognition.
Haven't run the code but the lines 580 and 581 seem to indicate that
https://github.com/muse1998/MMNet/blob/623e4b81d5de26d5354ffac1757c5af4afcc8d15/CASME2_3.py#L580-L581 There were data leakage problems since 19 years, so the first micro-expression study was also conducted to select the best results, and if we did not select the best, our results would be difficult to compare with existing results.
This work may suffer from a data leakage issues as in https://github.com/Justin900429/Facial-Separated-Micro-Expression-Recognition. Haven't run the code but the lines 580 and 581 seem to indicate that https://github.com/muse1998/MMNet/blob/623e4b81d5de26d5354ffac1757c5af4afcc8d15/CASME2_3.py#L580-L581
There were data leakage problems since 19 years, so the first micro-expression study was also conducted to select the best results, and if we did not select the best, our results would be difficult to compare with existing results.
Oh!thank you for your reply!
Hi, I tried to run CASMEII with your network, but I got overfitting under some LOSO, are you having similar problems?