desternylin / perfed Goto Github PK

Hello, when I was testing your code, about the projection matrix P for subspace training, I found that for example, when I tried to compress a 10000-dimensional model to 100, the shape of P is ( 100,10000 ), P * P ^ T is approximately equal to E, and P ^ T * P is not equal to E, which seems difficult to use for model compression ? Or is there a problem with the way I reproduce ?

Dear authors,

Thank you for your open-source code. It seems I cannot find the code to generate data partition for CIFAR10, FanshionMNIST, and EMNIST (only code for mnist and synthetic datasets are provided). I wonder whether you can upload the complete data partition code for the paper.

Best.