search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.55k stars 11.69k forks source link

what is suitable loss value while train? when should I stop it #1117

Open parquets opened 5 years ago

parquets commented 5 years ago

like the title, I set step 150 and finally I find the loss value is about 0.6, Is the value suitable?

stygian2a commented 5 years ago

The value depends entirely on your dataset. I seems good enough, but if your problem is supposed to be a very simple one, then it might not be so good. You should use tensorboard to see the evolution of the training and validation loss. As long as the validation is going down, you keep training. If at some point, it starts to go up while your training loss is still going down (or is staying the same), it means you are overfitting.

fastlater commented 5 years ago

Check this: https://github.com/matterport/Mask_RCNN/blob/master/assets/detection_tensorboard.png As you can see, the loss obtained when training the initial weights was really low so you can estimate that 0.6 is good but maybe not good enough to get art of state results. As @stygian2a mentioned, it will highly depends on your dataset, however, lower loss, better results.