Darknet: I train the network for 9000 iterations, If I add in more images, should I re-train from scratch or continue?

Created on 4 Aug 2020 · 4Comments · Source: AlexeyAB/darknet

I trained the network using a dataset after 9,000 iterations. I collected more images and added them into my dataset, should I re-train the network from scratch? or can I just keep training using the weight from 9000th iteration?

Source

sensationTI

Most helpful comment

If the classes were not changed, keep training with the saved weights. Your initial mAP differs from the previous dataset but it will be adjusting after some epochs with the new samples.

Yogerlan on 4 Aug 2020

👍2 😄1

All 4 comments

If the classes were not changed, keep training with the saved weights. Your initial mAP differs from the previous dataset but it will be adjusting after some epochs with the new samples.

Yogerlan on 4 Aug 2020

👍2 😄1

If the classes were not changed, keep training with the saved weights. Your initial mAP differs from the previous dataset but it will be adjusting after some epochs with the new samples.

If some objects in my newly collected images are not labeled, and I feed it into the network. Will it make any effect on the training result?

sensationTI on 4 Aug 2020

Yes. Images with no labels negative samples.

AnchorAnalytics on 5 Aug 2020

👍2

@sensationTI Note that images with desired objects without labels are not negative samples. They would impact your training and could confuse your model.