When DiffusonDet Model is created in 'train_net.py'?

[MatteoMele98]

Hi, first of all Congratulations on your work!

I was analyzing the code, starting from the 'train_net.py' file, to get a good understanding of how the different components worked together.

I can't figure out how the DiffusionDet model is actually created: in 'train_net.py' the build_model of d2 is invoked (as indicated here

DiffusionDet/train_net.py

Line 85 in 1efb36d

def build_model(cls, cfg):

), therefore the DiffusionDet class present in 'detectron_py' (which should create the DiffusionDet model) is not invoked (?)

I also do not understand when the forward and loss functions described in 'detectron_py' are called during training.

These are definitely trivial questions, but I can't figure them out. Thank you very much.

[MatteoMele98]

Looking closer I saw that build_model builds cfg.MODEL.META_ARCHITECTURE; using DiffusionDet configurations MODEL.META_ARCHITECTURE = DiffusionDet, so the model is created that way, right?

However, I still do not understand how the forward and loss function are invoked during training.

Thanks again