Re-train a custom network

I’m trying to add new logo detection to a custom network that I developed however I don’t know really how to re-train de network so that I can incorporate new logos without having to train it from zero.

I appreciate the help.

it might be possible to specify the initial training values in some software… but the simplest way is to just add the new images to the training set and re-run from zero. also, if you have new classes added you’ll pretty much have to i think.

If you now have too many images for the training software to handle, i chose to first run the new images through the existing model, and then i only add the new images that were incorrectly labeled (after i corrected them manually). results were pretty good. i had 1000s, so wrote a script to help with the checking, but this should be relatively easy to do manually as well.

note - if you’re using the google colab training, i think it can handle more images if you link them from google drive. i remember something like that.

1 Like