search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.53k stars 11.68k forks source link

overfitting problem #2382

Open Rimehdaoudi opened 3 years ago

Rimehdaoudi commented 3 years ago

hi am trying this implementation on my own dataset and while visualize the loss function curve a realize that there's an overfitting problem. also when i get a different mAP every time i run the model on the test set images( the test images are used only for test)! can anyone tell me what happen. here's the loss curve

image

pafechet commented 3 years ago

Hi @Rimehdaoudi How much images you have for training and validation ? How much class and are they well distributed ?

Rimehdaoudi commented 3 years ago

Hi @pafechet A total of 540 images were selected randomly and used for the model training with 80% for training and 20% for validation for the classes it's for brain tumor detection so we have tumor + background

pafechet commented 3 years ago

@Rimehdaoudi Ok you have quite a good amount of image for mask r-cnn, actually it's a transfert learning model so you could use even less. Well im sorry i don't really know right now. Check your bacth size too....

konstantin-frolov commented 3 years ago

You don't have overfitting problem. Overfitting is when val loss start increase after decreasing, but train loss сontinues to decrease.