bzgeng / iprls

Code and data for the SIGIR'2021 paper "Iterative Network Pruning with Uncertainty Regularization for Lifelong Sentiment Classification"

IPRLS code for Iterative Pruning with Regularization for Lifelong Sentiment Classification

requirements

• Python >=3.7
• Pytorch 1.2.0
• transformers

bert-base-uncased version BERT model need to be download from https://huggingface.co/bert-base-uncased , set it under path BERT/

You can run IPRLS with

$ bash experiment/run_IPRLS.sh 

After completing the above process, you need to run following bash to obtain final results

$ bash experiment/eval_middle_results.sh 

Run IPRLS with random task order

$ bash experiment/run_with_random_task_order.sh 

To evaluate shuffle order results

$ bash experiment/eval_shuffle_middle_results.sh 

iprls's Issues

Inference of the IPRLS

The inference stage of such kind of continual learning methods should be executed with the task-ID? For different tasks, the known ID will be used to activate task-related weights. And the number of classifiers may be different for different tasks. Am I right?