spider_dg

Domain Generalization via Meta-Learning

Prepare Data

(English) Spider

Download from Spider or using bash:

pip install gdown
gdown https://drive.google.com/uc\?export\=download\&id\=1_AckYkinAnhqmRQtGsQgUKAnTHxxX5J0
cd data/spider/
unzip spider.zip
mv spider raw

Chinese Spider

Download data and place it at data/ch_spider/raw.

Train BERT models for Spider

First, preprocess the data:

tensor2struct preprocess configs/spider/run_config/run_en_spider_bert_baseline.jsonnet

To train a supervised model, run

python experiments/spider_dg/run.py train configs/spider/run_config/run_en_spider_bert_baseline.jsonnet

To meta-train a supervised model, run

python experiments/spider_dg/run.py meta_train configs/spider/run_config/run_en_spider_bert_dgmaml.jsonnet

Evaluation

Run

tensor2struct eval config_file

where config_file can be the baseline config or bert_dgmaml config. For Chinese Spider, you need the following configs:

• configs/spider/run_config/run_ch_spider_bert_baseline.jsonnet
• configs/spider/run_config/run_ch_spider_bert_dgmaml.jsonnet