Official implementation of "Generalized Intent Discovery: Learning from Open World Dialogue System", COLING2022 main conference
You can use the ./run_pretrain.sh script for IND pretraining. For different settings, you need to slightly modify the parameters of the script.
Running pretraining on GID-SD (take GID-SD-60% as an example):
python main_pretrain.py \
--dataset GID-SD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 60 \
--batch_size 256 \
--gpus 1 \
--precision 16 \
--max_epochs 200 \
--num_labeled_classes 31 \
--num_unlabeled_classes 46 \
--comment 31_46 \
--offline \
Running pretraining on GID-MD (take GID-MD-60% as an example):
python main_pretrain.py \
--dataset GID-MD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 60 \
--batch_size 256 \
--gpus 1 \
--precision 16 \
--max_epochs 200 \
--num_labeled_classes 60 \
--num_unlabeled_classes 90 \
--comment 60_90 \
--offline \
Running pretraining on GID-CD (take GID-CD-60% as an example):
python main_pretrain.py \
--dataset GID-CD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 60 \
--batch_size 256 \
--gpus 1 \
--precision 16 \
--max_epochs 200 \
--num_labeled_classes 60 \
--num_unlabeled_classes 90 \
--comment 60_90 \
--offline \
You can use the ./run_discovery.sh script for end-to-end generalized intent discovery. For different settings, you need to slightly modify the parameters of the script.
Running generalized intent discovery on GID-SD (take GID-SD-20% as an example):
for s in 0
do
python main_discover.py \
--dataset GID-SD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 20 \
--batch_size 512 \
--gpus 1 \
--seed $s \
--max_epochs 200 \
--num_labeled_classes 62 \
--num_unlabeled_classes 15 \
--pretrained pretrain_checkpoints/pretrain-bert-base-uncased-GID-SD-62_15.cp \
--num_heads 4 \
--comment 62_15 \
--precision 16 \
--offline \
done
Running generalized intent discovery on GID-MD (take GID-MD-20% as an example):
for s in 0
do
python main_discover.py \
--dataset GID-MD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 20 \
--batch_size 512 \
--gpus 1 \
--seed $s \
--max_epochs 200 \
--num_labeled_classes 120 \
--num_unlabeled_classes 30 \
--pretrained pretrain_checkpoints/pretrain-bert-base-uncased-GID-MD-120_30.cp \
--num_heads 4 \
--comment 120_30 \
--precision 16 \
--offline \
done
Running generalized intent discovery on GID-CD (take GID-CD-20% as an example):
for s in 0
do
python main_discover.py \
--dataset GID-CD \
--data_dir dataset/GID_benchmark/ \
--OOD_ratio 20 \
--batch_size 512 \
--gpus 1 \
--seed $s \
--max_epochs 200 \
--num_labeled_classes 120 \
--num_unlabeled_classes 30 \
--pretrained pretrain_checkpoints/pretrain-bert-base-uncased-GID-CD-120_30.cp \
--num_heads 4 \
--comment 120_30 \
--precision 16 \
--offline \
done
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent