Flat-Lattice-Transformer
code for ACL 2020 paper: FLAT: Chinese NER Using Flat-Lattice Transformer.
Models and results can be found at our ACL 2020 paper FLAT: Chinese NER Using Flat-Lattice Transformer.
Requirement:
Python: 3.7.3
PyTorch: 1.2.0
FastNLP: 0.5.0
Numpy: 1.16.4
you can go here to know more about FastNLP.
How to run the code?
-
Download the character embeddings and word embeddings.
Character and Bigram embeddings (gigaword_chn.all.a2b.{'uni' or 'bi'}.ite50.vec) : Google Drive or Baidu Pan
Word(Lattice) embeddings:
yj, (ctb.50d.vec) : Google Drive or Baidu Pan
ls, (sgns.merge.word.bz2) : Baidu Pan
-
Modify the
paths.py
to add the pretrained embedding and the dataset -
Run following commands
python preprocess.py (add '--clip_msra' if you need to train FLAT on MSRA NER dataset)
cd V0 (without Bert) / V1 (with Bert)
python flat_main.py --dataset <dataset_name> (ontonotes, msra, weibo or resume)
If you want to record experiment result, you can use fitlog:
pip install fitlog
fitlog init V0
cd V0
fitlog log logs
then set use_fitlog = True in flat_main.py.
you can go here to know more about Fitlog.
Cite:
运行环境:
Python: 3.7.3
PyTorch: 1.2.0
FastNLP: 0.5.0
Numpy: 1.16.4
你可以在 这里 深入了解 FastNLP 这个库.
如何运行?
-
请下载预训练的embedding
从Google Drive 或 Baidu Pan 下载字和 Bigram 的 embedding (gigaword_chn.all.a2b.{'uni' or 'bi'}.ite50.vec)
从Google Drive 或 Baidu Pan 下载词的 embedding (ctb.50d.vec)(yj)
从Baidu Pan 下载词的embedding (sgns.merge.bigram.bz2)(ls)
-
修改
paths.py
来添加预训练的 embedding 和你的数据集 -
运行下面的代码
python preprocess.py (add '--clip_msra' if you need to train FLAT on MSRA NER dataset)
cd V0 (without Bert) / V1 (with Bert)
python flat_main.py --dataset <dataset_name> (ontonotes, msra, weibo or resume)
如果你想方便地记录和观察实验结果, 你可以使用fitlog:
pip install fitlog
fitlog init V0
cd V0
fitlog log logs
然后把flat_main.py里的 use_fitlog 设置为 True 就行 你可以在 这里 深入了解 Fitlog 这个工具
引用:
更新说明:
5.7共提交两个版本,其中V2使用tensor.unique()用于去除相对位置中重复组合(记为Flat_unique),V3使用标量替代了FLAt中的相对位置编码(记为Flat_scalar).详见FLAT瘦身日记
使用这两种方法的显存占用如下表所示
batch_size=10
seq_len | 50 | 100 | 150 | 200 | 250 | 300 |
---|---|---|---|---|---|---|
Flat | 1096MB | 1668MB | 2734MB | 4118MB | 5938MB | 8374MB |
Flat_unique | 964MB | 1204MB | 1610MB | 2166MB | 2922MB | 3940MB |
Flat_scalar | 878MB | 916MB | 1028MB | 1062MB | 1148MB | 1322MB |
Bert+Flat | 1605MB | 2237MB | 3333MB | 4725MB | 6571MB | 9039MB |
Bert+Flat_unique | 1495MB | 1685MB | 2129MB | 2697MB | 3453MB | 4585MB |
Bert+Flat_scalar | 1409MB | 1481MB | 1565MB | 1617MB | 1705MB | 2051MB |