Validation

[nasserha]

Hello,

I see that the detectnet v2 documentation has an validation_data_source option in the dataset config. I wonder whether it’s possible to use this option also for the fasterrcnn?

Otherwise, the fasterrcnn has a validation_fold option. When this option is set to zero, one can still see an mAP value coming out.

• Does this mean that the mAP is measured on the training set?
• If yes, when set to 1, what happens, does it measure on the validation set if there is a val_split provisionned in the tfrecords data preparation?

Thanks

[Morganh]

@nasserha

I wonder whether it’s possible to use this option also for the fasterrcnn?

Yes, you can use it.

@nasserha

When this option is set to zero, one can still see an mAP value coming out.

• Does this mean that the mAP is measured on the training set?

When validation_fold is set to zero, the mAP is measured on the validation set.
Refer to DetectNet_v2 - NVIDIA Docs. “This partition is available in fold 0 of the TFrecords generated.”

[nasserha]

Thanks @Morganh. I appreciate your help.