The code for CCF-BDCI-Sentiment-Analysis-Baseline
1.从该开源代码中改写的
2.该模型将文本截成k段,分别输入语言模型,然后顶层用GRU拼接起来。好处在于设置小的max_length和更大的k来降低显存占用,因为显存占用是关于长度平方级增长的,而关于k是线性增长的
模型 | 线上F1 |
---|---|
Bert-base | 80.3 |
Bert-wwm-ext | 80.5 |
XLNet-base | 79.25 |
XLNet-mid | 79.6 |
XLNet-large | – |
Roberta-mid | 80.5 |
Roberta-large (max_seq_length=512, split_num=1) | 81.25 |
注:
1)实际长度 = max_seq_length * split_num
2)实际batch size 大小= per_gpu_train_batch_size * numbers of gpu
3)上面的结果所使用的是4卡GPU,因此batch size为4。如果只有1卡的话,那么per_gpu_train_batch_size应设为4, max_length设置小一些。
4)如果显存太小,可以设置gradient_accumulation_steps参数,比如gradient_accumulation_steps=2,batch size=4,那么就会运行2次,每次batch size为2,累计梯度后更新,等价于batch size=4,但速度会慢两倍。而且迭代次数也要相应提高两倍,即train_steps设为10000
具体batch size可看运行时的log,如:
09/06/2019 21:03:41 - INFO - main - ***** Running training *****
09/06/2019 21:03:41 - INFO - main - Num examples = 5872
09/06/2019 21:03:41 - INFO - main - Batch size = 4
09/06/2019 21:03:41 - INFO - main - Num steps = 5000
请查看该网站了解赛题
从该网站中下载数据集, 并解压在./data目录。
cd data
python preprocess.py
cd ..
bash run_bert.sh
#5 fold取平均
python combine.py --model_prefix ./model_bert --out_path ./sub.csv
从该网站下载pytorch权重,并解压到chinese_wwm_ex_bert目录下: https://github.com/ymcui/Chinese-BERT-wwm
bash run_bert_wwm_ext.sh
python combine.py --model_prefix ./model_bert_wwm_ext --out_path ./sub.csv
从该网站下载pytorch权重,并解压到./chinese_xlnet_mid/目录下: https://github.com/ymcui/Chinese-PreTrained-XLNet
bash run_xlnet.sh
python combine.py --model_prefix ./model_xlnet --out_path ./sub.csv
从该网站下载tensorflow版本的权重,并解压到./chinese_roberta/目录下: https://github.com/brightmart/roberta_zh
mv chinese_roberta/bert_config_middle.json chinese_roberta/config.json
python -u -m pytorch_transformers.convert_tf_checkpoint_to_pytorch --tf_checkpoint_path chinese_roberta/ --bert_config_file chinese_roberta/config.json --pytorch_dump_path chinese_roberta/pytorch_model.bin
bash run_roberta.sh
python combine.py --model_prefix ./model_roberta --out_path ./sub.csv