The.pth model saved by save_checkpoint is inconsistent with the.pth model exported by this model about aimet HOT 5 OPEN

What are the differences are you seeing?
Did you use prepare model API?

from aimet.

I printed the output results of both and the results were completely inconsistent. I did not use the prepare model API

from aimet.

If you are performing QAT, then weights are expected to change (in order to reduce the quantization noise) which is reflected in the exported model.

from aimet.

My QAT fine is over, the optimal.pth model saved in the finetune process, and the pth exported by the optimal.pth model export should be consistent. I don't understand what's new here that's causing the model to turn out differently

from aimet.

If you look at the model saved by save_checkpoint it is a quantsim model whereas .export exports out a model without Quantization wrappers.

from aimet.