UV decomposition

[icey-zhang]

Hello, I would like to ask which part of the implementation of UV decomposition mentioned in the paper, I did not find the code of this part

[shupenggui]

Please take a look at this line https://github.com/VITA-Group/ATMC/blob/ecf61a6f07525c8e259db764513c0820259be319/super_module/super_class.py#L485

For the implementation, the DeepLRModel could substitute the sp one as the super class for models like ResNet34.

[icey-zhang]

But the code use DeepSP_v2Model rather than DeepLRModel (cidar10, ResNet34), Is there any difference between them? Do the weightA and weightB correspond to the decomposition values? what do the modelu and modelz stand for?

[shupenggui]

But the code use DeepSP_v2Model rather than DeepLRModel (cidar10, ResNet34), Is there any difference between them?

Please refer to this line https://github.com/VITA-Group/ATMC/blob/ecf61a6f07525c8e259db764513c0820259be319/super_module/super_class.py#L793 if you are trying to figure out how the decomposition work in ATMC.

By the way, the DeepLRModel serves at the Low-rank baseline in the paper. I thought you were trying to reproduce that one in your previous question. No relation exists between those two super modules.

Do the weightA and weightB correspond to the decomposition values?

Yes, those two weights are the decomposition matrices for U and V.

what do the modelu and modelz stand for?

Those models refer to the duplicated optimization variables applied in ADMM.