A-Klass/torch_topkast

Issues

• insert weight rescaling after sparsification

  #10 opened 3 years ago by lisa-wm
  0

• Loss regularization doesn't work for B\A

  #18 opened 3 years ago by SvenLorenz
  1

• make backward pass truly sparse by sparsifying the outer product computation

  #12 opened 3 years ago by lisa-wm
  1

• make layer attributes for sparse weights more efficient

  #11 opened 3 years ago by lisa-wm
  1

• restrain changes of active set (and thus coo initialization) to every m steps

  #13 opened 3 years ago by lisa-wm
  0

• CSR is apparently faster and less storage intensive than COO

  #9 opened 3 years ago by SvenLorenz
  2

• Having permanent sparse active weight matrix is maybe faster

  #8 opened 3 years ago by SvenLorenz
  1