liuyuqi-dellpc
f36d22ed8d
增加 notebook
|
5 years ago | |
|---|---|---|
| .vscode | 5 years ago | |
| data | 5 years ago | |
| img | 5 years ago | |
| saved_model | 5 years ago | |
| sqlnet | 5 years ago | |
| .dockerignore | 5 years ago | |
| .gitignore | 5 years ago | |
| Dockerfile | 5 years ago | |
| LICENSE | 5 years ago | |
| README.cn.md | 5 years ago | |
| README.md | 5 years ago | |
| colab-nl2sql.ipynb | 5 years ago | |
| docker-compose.debug.yml | 5 years ago | |
| docker-compose.yml | 5 years ago | |
| pip.conf | 5 years ago | |
| requirements.txt | 5 years ago | |
| start_test.sh | 5 years ago | |
| start_train.sh | 5 years ago | |
| test.py | 5 years ago | |
| train.py | 5 years ago |
README.cn.md
nl2sql 文字问题转 sql
https://drive.google.com/open?id=10RtAom_D4zOp_w5OsYLTtLd2TEuC7C1X
模型训练所需要的算力还是挺大的，在 K80 上训练一个 Epoch 大概需要 7 到 8 分钟，基线模型默认训练 100 个 Epoch。追一表示试验多次后，验证集上的平均 Logic Form Accuracy 为 27.5%，即 SQL 每一个子句全都预测正确的概率为 27.5%。
WikiSQL 基准：https://github.com/salesforce/WikiSQL
CUDA_VISIBLE_DEVICES=0 python train.py --ca --gpu --bs 128