This project is WIP.
pip install dlk
or
git clone this repo and do
python setup.py install
- Provide a templete for deep learning (especially for nlp) training and deploy.
- Provide parameters search.
- Provide basic architecture search.
- Provide some basic modules and models.
- Provide reuse the pretrained model for predict.
-
Tasks support
-
NLP
- Classification
- Pair Classification
- Regression
- Pair Regression
- Sequence Labeling
- Span Classification
- Relation Extraction
- Token Rerank
- MRC SQuAD
- Translation
- Summary
-
CV
- Classification
-
NLP
-
Generate models.
-
Distill structure.
-
Ensemble models for NLU(and check how to do this in NLG)
-
Training Strategy
- Adversarial Training(FGM/PGD/FreeLB)
- Schedule Loss(you can control the loss schedule)
- Schedule MultiTask Loss(you can control the loss schedule for each task)
- Focal Loss
-
Online service by triton.
-
Data Augment.
-
Support LightGBM. Will split to another package.
-
Make most complexity modules like Beam Search, CRF to be scriptable.
-
Add UnitTest
- Parser
- Tokenizer
- Config
- Link