I have trained the base model using MMdetection starter notebook and I need some clarification here. I don’t know why I am getting lower mAP on test data than validation dataset. Model’s validation segm mAP is gradually increasing after each epoch but it is almost same mAP on test dataset even though the data pipeline is same for both datasets. Can anybody explain why it is overfitting? or Is there any hyperparameter that I am missing here?
Thanks in advance.
2 Likes
Hi @saidinesh_pola,
The validation set in this release is just for integration testing.
It contains images from the training dataset due to which your model might be getting higher mAP on validation set v/s the actual test data. This is different from our previous dataset releases, and we should have communicated it better. Sorry for any confusion caused.
We suggest you to create your validation set by splitting the training dataset for your local testing.
Yes, in this round val dataset has been copied to train dataset. So, you should just remove images that are in val dataset from the train dataset by yourself, and you will get same workflow as in round 1.
Hope it will help you.
3 Likes