Closed charleslparker closed 3 years ago
unmonoqueteclea
commented
3 years ago Just for curiosity, two more questions:
unmonoqueteclea
commented
3 years ago 
charleslparker
commented
3 years ago Are new serialized models heavier than previous ones?
No, the JSON should be about the same. All of the conversion formats (tflite, smbundle, etc.) are significantly lighter weight.
Will these new changes make it easier to start training from a previously trained model?
No, that was already fairly easy, I just haven't done it yet. I'm updating pyramid with the new sensenet as we speak, so I'll take the chance to do it now.
Do I need to do something to invalidate legacy sensenet_metada.json file as it happened in previous versions?
No. I've started including this file in the package so it will always be up-to-date.
unmonoqueteclea
commented
3 years ago Great, thank you very much for all your answers, shall I merge?
charleslparker
commented
3 years ago Great, thank you very much for all your answers, shall I merge?
No, I will do it. Thanks for your questions!
Among other things, this PR introduces the full YOLOv4 topology. I took the chance to refactor how layers are serialized and deserialized to JSON, so that the format is now much more flexible and more closely matches to what TF/Pytorch/Keras do, so that new things will be easier to admit in the future.
There are a number of changes that will break clients here, and old object detectors are not compatible with this (yet). @unmonoqueteclea : How hard will it be to retrain models for the license plate stuff (once pyramid is using this version)? If it's a huge pain, I can probably write some code that will ingest the old format as well, but I'd like to avoid it if I can.
Pyramid needs some minor tweaking to be compatable with this.