Open rotem154154 opened 2 years ago
Hi, Sorry for the late reply. Empirically, we found that resetting part of the generator helps with the diversity of the synthesized samples in the end and (slightly) improves the performance of distillation. The bug (AttributeError:...) has been fixed. Thanks for pointing it out.
thanks so much for open sourcing your work!
when i train the student from the datafree_kd.py with the parameters in README.md the network is training untill epoch 138 when the if in fast_meta.py line 110 :
can you explain why reseting the network and why the code crush with the error: AttributeError: 'Generator' object has no attribute 'weight' incide the reset_l0 function: https://github.com/zju-vipa/Fast-Datafree/blob/0955a6cabb60ddb4bf1943d2957bc2ad011f3155/datafree/synthesis/fast_meta.py#L33