Skip to content

Latest commit

 

History

History
6 lines (4 loc) · 512 Bytes

File metadata and controls

6 lines (4 loc) · 512 Bytes

Weighted-low-rank-factorization-Pytorch

PyTorch implementation of the paper Language model compression with weighted low-rank factorization (https://arxiv.org/abs/2207.00112 ICLR 2022)

Open source implementation of FWSVD(Factorized weighted Singular Value decompisition) training method introduced in https://arxiv.org/abs/2207.00112 using pytorch, accelerate & HuggingFace

Run train_glue_fwsvd.sh to run a pre-trained BERT 12 layer model on GLUE tasks with FWSVD. Use TASK_NAME flag to change the task