I'm working on the multi-classification task (6 classes in total), and got almost perfect training and test accuracy (over 99%) based on my trained cnn model. I am trying to know that if my trained model suffers from overfitting or not. Attached are the learning curves of loss and accuracy for that model. The whole training dataset contains roughly 30k samples and the ratio of training to validation dataset is 4:1. Does my loss curve show overfitting?
CodePudding user response：
Can I ask how many label do you need to classify? From the charts, seem to be no sign of overfitting with 30k samples.
However, if the samples in your dataset is really similar, which results in your training and validation set is similar, than your model is probably overfit to this specific dataset.
If real-world data actually looks like your dataset, than nothing to be worried about. However, if not, I recommend to expand the dataset in term of variety to better generalize real-world data.
CodePudding user response：
First, training accuracy is mostly meaningless (so ignore that).
The symptom of over-fitting that you could see with train/val/test accuracy/loss curves is:
- Low Training Loss / high training accuracy
- and Low (decreasing) validation accuracy (increasing validation loss).
Those are symptoms of over-fitting because: you have learned from the training data (training loss became small, training accuracy became good) but that learning doesn't transfer to the validation data.
I don't see that in your curves, so it doesn't appear to be over-fitting.