srush / annotated-s4

Implementation of https://srush.github.io/annotated-s4
https://srush.github.io/annotated-s4
MIT License
460 stars 61 forks source link

Add LRA Listops and IMDB experiments #53

Closed jxiw closed 2 years ago

jxiw commented 2 years ago

ListOps

Downolad dataset in GitHub page. Unzip the downloaded dataset, and move it to the project folder.

python -m s4.train --dataset listops-classification --model s4 --epochs 100 --bsz 50 --d_model 128 --n_layers 4 --ssm_n 64 --lr 8e-3 --p_dropout 0

Training with the previously defined hyper-parameters yields the test accuracy 50.2% on the test set.

Screen Shot 2022-04-08 at 10 43 04 AM

IMDB

We used the huggingface datasets for LRA IMDB review task.

python -m s4.train --dataset imdb-classification --model s4 --epochs 30 --bsz 50 --d_model 64 --n_layers 4 --ssm_n 64 --lr 1e-2 --p_dropout 0

Training with the previously defined hyper-parameters yields the test accuracy 78.42% on the test set.

Screen Shot 2022-04-07 at 2 03 57 AM
srush commented 2 years ago

Need to run the linter. make autoformat

srush commented 2 years ago

Actually this is not a real error. Turned it off in the linter.