thu-spmi / PPT2DST

Prompt Pool based Class-Incremental Continual Learning for Dialog State Tracking

This is the official code for "Prompt Pool based Class-Incremental Continual Learning for Dialog State Tracking" (ASRU2023). This code is based on CPT4DST. The package requirements and the dataset processing is the same as CPT4DST.

Experiments

The following script will sequentially execute model training and evaluation.

python prompt_pool_tuning.py \
    --train data/sgd_train.json \
    --dev data/sgd_dev.json \
    --test data/sgd_test.json \
    --schema data/all_schema.json \
    --select_method bert-encoder \
    --epochs 3\
    --learning_rate 2e-5 \
    --batch_size 8 \
    --gpu_id 0 \
    --dataset_order 1 \ # The order of training tasks
    --top_n 10 \ # the number of selected prompts for each task
    --pool_size 150\ # the number of prompts in the prompt pool
    --penalty \ # whether to add penalty items
    --M 50 \ # the rehearsal buffer size
    --dis_method euclidean\ # the method of calculating distance
    --model_name t5-small \ # the backbone, can be chosen from 'google/mt5-small', 't5-small', 't5-base', 't5-large'
    --dataset sgd \ # the dataset, sgd or CM-Pickup (not released currently)
    --memory_type fix_size \ # the storage method for rehearsal buffer, fix_size (maintain a fixed buffer) or incre_size (store the same number of samples for each task)

• Please contact [e-mail] if you have any questions.