yphacker / text_classification_pytorch

Toxic Comment Classification Challenge

竞赛链接

数据下载

data

评估标准

the score is the average of the individual AUCs of each predicted column

kaggle score:

5折

model	offline score	online score	note
ml	-1	0.97016
cnn	-1	0.96977
cnn+预训练	-1	0.97571
rnn	-1	0.95663
rnn+atten	-1	0.97542
rcnn	-1	0.97468
bert(bert-base-uncased)	0.9900	0.98563	0.9895,0.9912,0.9899,0.9888,0.9908
albert(albert-base-v2)	0.9720	-1	0.9686,0.9741,0.9714,0.9730,0.9727
xlmroberta(xlm-roberta-base)
bart(bart-large-cnn)

单折 |model|offline score|note| |:---:|:---:|:---:|:---:| |ml|-1|| |cnn|-1|| |cnn+预训练|-1|| |rnn|-1|| |rnn+atten|-1|| |rcnn|-1|| |bert(bert-base-uncased)|-1|| |albert(albert-base-v2)|98.22|epoch=5就不再提升了| |xlmroberta(xlm-roberta-base)|||| |bart(bart-large-cnn)||||

实验环境

Tesla P100 16G cuda9
python:3.6
torch:1.2.0.dev20190722

script

nohup python main.py -m='cnn' -b=256 -e=3 > nohup/cnn.out 2>&1 &
nohup python main.py -m='bert' -b=32 -e=4 > nohup/bert.out 2>&1 &
nohup python main.py -m='albert' -b=64 -e=8 -mode=2 > nohup/albert.out 2>&1 & nohup python main.py -m='albert' -b=64 -e=5 > nohup/albert.out 2>&1 &
nohup python main.py -m='xlmroberta' -b=20 -e=2 > nohup/xlmroberta.out 2>&1 &
python predict.py -m='bert'

参考文章

参考代码

[1] google-research/bert
[2] google-research/ALBERT
[3] huggingface/transformers
[4] 649453932/Bert-Chinese-Text-Classification-Pytorch