/dl_active_learning_project

Project for Deep Learning course: Deep neural models for active learning in the sequence tagging tasks

Primary LanguagePython

Deep Learning course project on Active Learning

Deep neural models for active learning in the sequence tagging tasks

Repository structure

  • data -- a place with training and evaluation datasets on which we conducted our experiments
  • experiments -- scores, that we got after experimenting with different models and strategies. Each result folder consists of several runs in order to provide the variance estimation
  • src -- main scripts to run experiments on your own
  • configs -- main files with paramethers to run the experiments

Running example

Suppose you want to run an Active Learning emulation using BERT with Monte-Carlo Dropout extension

To run the experiment in our setting you should write the config pass and the name of our main script, as following:

HYDRA_CONFIG_PATH='../configs/al_exp_bert.yaml' python run_active_learning.py

Also, you can determine the desired cuda device (not the 0 one) before the config path:

CUDA_VISIBLE_DEVICES=2 HYDRA_CONFIG_PATH='../configs/al_exp_bert.yaml' python run_active_learning.py

You can enable multiprocessing via run_al_multiprocess.py. It works in the same way as run_active_learning.py. It will consume more RAM, please use it on your own risk!

Scripts assume that you do not already have statistics for the experiment in the experiments directory.

Keep in mind, that computations on large datasets like CoNLL-2003 can be rather heavy and assume that you are running program on gpu and have 6--8 GB of VRAM to spare.

PS We ran the experiments using clasters with several Nvidia GeForce 2080 and Tesla P40. 10GB of VRAM should be more than enough for a single run for the most computationaly expensive part.

Collaborators: Denis Belyakov, Lyubov Kuprianova, Miron Kuznetsov, Dmitri Puzyrev, Angelina Yaroshenko

Special thanks to our research supervisor Artem Shelmanov!