search

yanxp / MetaR-CNN

Meta R-CNN : Towards General Solver for Instance-level Low-shot Learning
https://yanxp.github.io/metarcnn.html
177 stars 23 forks source link

The accuracy of the model was 0!!!!! #58

Open qindengda opened 3 years ago

qindengda commented 3 years ago

Hello! When I train the model with the default settings,the training accuracy in the first phase is 0. I don't think the model has learned anything at all ,and I don't know the problem.

hanhao2019102935 commented 3 years ago

@1696427331 Have you solved the problem?Is it a data set construction problem？

qindengda commented 3 years ago

@1696427331 Have you solved the problem?Is it a data set construction problem？

I haven't solved this problem yet. I use the same data set and data set structure as mentioned in readme. There is no error in network training, and loss seems to be normal, but the accuracy is not right.

Jaraxxus-Me commented 3 years ago

@1696427331 Have you solved the problem?Is it a data set construction problem？

I haven't solved this problem yet. I use the same data set and data set structure as mentioned in readme. There is no error in network training, and loss seems to be normal, but the accuracy is not right.

Hi, have you solved the problem? I met the same problem. The mAP is 0. Besides, I also tried FSDet, which is based on this repo and met 0 mAP problem again. Everything seems normal, but the model can't produce anything. I noticed the training loss is about 1.5 from the forth epoch, which is much higher than the authors provided 0.3. And the meta loss strangely remains the same... What about your case? Thanks

qindengda commented 3 years ago

@1696427331 Have you solved the problem?Is it a data set construction problem？

I haven't solved this problem yet. I use the same data set and data set structure as mentioned in readme. There is no error in network training, and loss seems to be normal, but the accuracy is not right.

Hi, have you solved the problem? I met the same problem. The mAP is 0. Besides, I also tried FSDet, which is based on this repo and met 0 mAP problem again. Everything seems normal, but the model can't produce anything. I noticed the training loss is about 1.5 from the forth epoch, which is much higher than the authors provided 0.3. And the meta loss strangely remains the same... What about your case? Thanks

Sorry，I don't know the cause of the problem at the moment. The problem I encountered was the same as you. The training loss of the model was maintained at about 1.5, and the meta loss basically did not decrease.