search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.7k stars 11.71k forks source link

Suffer from over fitting during validation!! Will dropout help in reducing val. losses and over-fitting ? #1796

Open rakehsaleem opened 5 years ago

rakehsaleem commented 5 years ago

Hello, I have been looking to find a way to reduce overfitting for my model. Let me tell you that my training losses are converging but my validation losses are overfitting a big time and I have a good dataset of 1100 images with some augmentations applied. Since my objects are really thin like cracks so I need to remove overfitting. I know dropout can be helpful in reducing overfitting problem and I have looked at your previous model resources but found dropout only in fpn_classifier_graph as mentioned in above comments.

I am not sure will adding dropout to this will solve the issue or do I have to add it in the conv. block of ROI pooling. Similarly, do I have to add it in the RPN network and Resnet graph also? Any help would be really appreciated. @waleedka

sohinimallick commented 3 years ago

@rakehsaleem did you solve this?

rakehsaleem commented 3 years ago

@sohinimallick Thanks for reaching out, I have done this project in the past (at least more than a year ago) and at that time, I only tried adding an RPN network and to be honest, the results were nonetheless the same and I don't see much of a difference in validation loss. I could say maybe it's because of the reason I was working on crack damages (object with very fine line and random structure). I would like to know if you achieve some good results by adding this? Please share your findings here so others can know anything critical.

Eventually, I tried multi-stage training and keep a check on my training set up to stop when a sufficient level or loss is achieved. Hope this helps and please ask me any question if you have! Thanks :)