I have successfully retrained the entry-level classification of V3 using the Python2.7 api for my own 100 classes and it gives decent results, but not exceptionally well.
I also have code to redirect the entire network from scratch, as indicated here (google code) , but this is a resource and time, and I have 400,000 images, so I donβt know what accuracy will be after training.
I was wondering if I can reinstall some of the last few fully related layers or more than just a classification layer so that the accuracy can be improved to some degree, and it is also not very difficult in terms of resources and time.
I tried to search a lot, but could not find anything. Perhaps what I want to do? and I need help with this.
source share