Skip to main content
fastai_community
Projects
glue-benchmark
Reports
LR-finder approaches face-off

LR-finder approaches face-off

Comparison of effectiveness of LR-finder suggestions for NLP tasks on GLUE benchmark
Arto
Created on May 5|Last edited on May 10

SST2

Binary classification task. Dataset size: train - 67k, valid - 872. Fine-tuning whole model (no "freezing").
The results are averaged across 5 independent lr_find() calls. To explore consistency over runs one can ungroup the results for certain method and compare individual runs.

group: slide-glue-sst2-distilroberta-basegroup: valley-glue-sst2-distilroberta-basegroup: steep-glue-sst2-distilroberta-basegroup: minimum-glue-sst2-distilroberta-base0.00.10.20.30.40.50.60.70.80.9
3k4k5k6k7k8kStep00.20.40.60.81
group: slide-glue-sst2-distilroberta-base
group: valley-glue-sst2-distilroberta-base
group: steep-glue-sst2-distilroberta-base
group: minimum-glue-sst2-distilroberta-base
Run set
20


CoLA

Binary classification task. Dataset size: train - 8.5k, valid - 1k.

Run set
20


MRPC

Binary classification task. Dataset size: train - 3668, valid - 408.

Run set
20


Created with ❤️ on Weights & Biases.

https://wandb.ai/fastai_community/glue-benchmark/reports/LR-finder-approaches-face-off--Vmlldzo2NjM3NTA
Made with Weights & Biases. Sign up or log in to create reports like this one.