Add new images into dataset after pruning and retraining.

Hello Morganh,

I had one problem that is I have trained model for a class with some dataset and pruned it and retrained it but result was not good so I had increased data size. can I add new data in the previous dataset generate tf-records and start from retraining ?? It will give me good result or I will have to start from training again. ?

I think it can.
But if your training result is quite poor, the tlt model may not be a good option to work as pre-trained model during retraining.

Thanks Morganh.
If I will get good result after retraining then I can go for retraining again using retrained weight for new images of the class and can improve model further right ?? But will not be able to add new class in the retraining. right??

Sure.
You can retrain many times.
You can compare training spec and retraining spec for better understanding.
The most difference is the changing of pre-trained model.

If you add new class in the retraining, but your prior trained tlt model does not cover the new class, so you can consider your retraining as the first training for the new class.

Thanks Morganh,
I did this process. I train model and retrain it and use the retrained model weights resnet10_detector.tlt as base weights for re-training second. In re-training 2 I remove all the data for image_2 and label_2 and add new data for the new class and some data of the previous class so the result which I get is poor but in training 1 it was good. so can you suggest how should I go so that I can get a good result?

Thanks.

It is better for you to generate the tfrecords for all the target classes and then trigger training.

yes, I had also done this before re-training 2 then start re-training.

Also, in your latest re-training, please use the original pre-trained model in ngc.