This repository is the official implementation of UCBEE The experimental section of UCBEE can be divided into two major parts:
Part 1 (Finetuning and predictions): We finetune the ElasticBERT backbone after attaching exits to all the layers on RTE, SST-2, MNLI and MRPC (GLUE) datasets and then obtain prediction as well as confidence values for the evaluation (SciTail, IMDB, Yelp, SNLI, QQP)(GLUE and ELUE datasets except IMDB) i.e. all exit predictions for all samples (num_samples X num_exits)
Part 2: After training of exits, save the predictions and confidence of each exit point as given in (Early_Exit_Confidence_data_yelp____.csv).
For UCBEE with one exit execute the "ubert one exit.ipynb" file.
To install requirements
pip install -r requirements.txt
GLUE datasets are available at: GLUE Datasets
ELUE datasets are available at: Elue Datasets
Yelp dataset could be found here: Yelp dataset
IMDb dataset could be created by running "Create_IMDb_tsv_files.ipynb" code.
Finetuning and pre-training the multi-exit model (Part-1) is closely based on ElasticBERT, we acknowledge and thank the authors for the codebases.