Skip to content

Latest commit

 

History

History
7 lines (6 loc) · 424 Bytes

README.md

File metadata and controls

7 lines (6 loc) · 424 Bytes

Sparse_Sketch_Reducer

We implement the sparse sketch reducer (S2 Reducer) on pytorch 1.9. Folder LSTM/ contains the LSTM training code, which is a custom training code based on 'PowerSGD'. Folder GNMT/ contains the GNMT-8/16 training code, which is a custom training code based on the benchmark of MLPref. These two models yield sparse gradient from embedding matrix. ./codings contains the main code of sparse-sketch.