Closed maximzubkov closed 3 years ago
I've also experienced this. I tried training on the ListOps task multiple times, and sometimes I get an accuracy around 16, sometimes around 27.
See also #6, #14, #20.
@cifkao Thank you!
This has been fixed in the recent update. Please try again and let us know if you have any trouble. Sorry for the delay. These emails have been going to my spam folder.
Hello again!
Thank you for your work and for open-sourcing the codebase! I tried to run a listops experiment with
Linear transformer
and got results on the test that did not correspond to results proposed in the paper:The model config was absolutely identical to the one used in
transformer
and the only thing I changed in thelra_benchmark/listops/train.py
, was the following lines:The experiments were run inside a docker
nvidia/cuda:11.0-cudnn8-devel-ubuntu18.04
on 4 Tesla T4 with requirements:and
I can share the full
Dockerfile
if needed