Awesome

ConPET

Source code for ConPET: Continual Parameter-Efficient Tuning for Large Language Models.

Prepare the Datasets

Download the data from Google Drive link. Put the files in the sub-folder with name data. You should get the following structure:

ConPET/
|-- data
    |-- fewnerd
    |-- fewrel

Note that the other 4 datasets (OntoNotes, BBN, TACRED, ACE 2005) are not available due to the license issue.

To prepare for CRL, run the following commands in advance:

ln -s ../data CRL/data
ln -s ../scripts CRL/scripts

Prepare the Environment

You can get all the information of our Conda experimental environment in env.yaml.

Meaning of Arguments

The following arguments will be frequently mentioned in commands:

<dataset_name>: name of the dataset, chosen from "fewnerd", "ontonotes", "bbn", "fewrel", "tacred", and "ace"
<save_suffix>: a string that determines the suffix of your saving path of checkpoints and results
<seed>: an integer representing the random seed, we use 100, 200, and 300 in our repeated experiments
<device>: a string representing the device, such as "cuda:0"
<infer_device>: a string representing a device for inference, generally different from <device>

Results of Baselines

For EMR, run the following command:

python3 train_cycle.py --dataset_name <dataset_name> --use_selector 0 --batch_limit_policy 0 \
  --cycle_suffix <save_suffix> --method_type linear --continual_method emr \
  --seed <seed> --start 1 --device <device> --clear

For EA-EMR, run the following command:

python3 train_eaemr.py --dataset_name <dataset_name> --start 1 --cycle_suffix <save_suffix> \
  --batch_limit_policy 0 --seed <seed> --start 1 --device <device> --clear

For CRL, run the following commands:

cd CRL
python run_continual.py --device <device> --dataset <dataset_name> \
  --log_path <save_suffix> --sample_k 500 --seed <seed>
cd ..

For all the above three methods, you should simultaneously run a parallel process of the following command to select memorized samples:

python3 kmeans_client.py --dataset_name <dataset_name> --method_type linear \
  --cycle_suffix <save_suffix> --local --start 1

Results of Static ConPET

EMR* (Static ConPET + EMR)

Run the following command:

python3 train_cycle.py --dataset_name <dataset_name> --use_selector 0 --batch_limit_policy 2 \
  --cycle_suffix <save_suffix> --method_type linear --continual_method emr \
  --seed <seed> --start 1 --device <device> --clear

EA-EMR* (Static ConPET + EA-EMR)

Run the following command:

python3 train_eaemr.py --dataset_name <dataset_name> --start 1 --cycle_suffix <save_suffix> \
  --batch_limit_policy 2 --seed <seed> --device <device> --clear

CRL* (Static ConPET + CRL)

Run the following commands:

cd CRL
python run_continual.py --device <device> --dataset <dataset_name> \
   --log_path <save_suffix> --sample_k 500 --seed <seed> --dynamic_sampling
cd ..

Results of Dynamic ConPET

Run the following command:

python3 train_cycle.py --dataset_name <dataset_name> --use_selector 2 \
  --teacher_forcing --batch_limit_policy 2 \
  --cycle_suffix <save_suffix> --method_type linear --continual_method our \
  --seed <seed> --start 1 --device <device> --infer_device <infer_device>

Results of Limitless

Run the following command:

python3 train_cycle.py --dataset_name <dataset_name> --use_selector 2 \
  --teacher_forcing --batch_limit_policy 0 \
  --cycle_suffix <save_suffix> --method_type linear --continual_method our \
  --seed <seed> --start 1 --device <device> --infer_device <infer_device>