# CCF-BDCI-Sentiment-Analysis-Baseline **Repository Path**: singleCH/CCF-BDCI-Sentiment-Analysis-Baseline ## Basic Information - **Project Name**: CCF-BDCI-Sentiment-Analysis-Baseline - **Description**: The code for CCF-BDCI-Sentiment-Analysis-Baseline - **Primary Language**: Unknown - **License**: Apache-2.0 - **Default Branch**: master - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 0 - **Forks**: 0 - **Created**: 2020-04-01 - **Last Updated**: 2020-12-19 ## Categories & Tags **Categories**: Uncategorized **Tags**: None ## README # CCF-BDCI-Sentiment-Analysis-Baseline 1.从该[开源代码](https://github.com/huggingface/pytorch-transformers)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 ## 赛题说明 请查看该[网站](https://www.datafountain.cn/competitions/350)了解赛题 ## 下载数据集 从该[网站](https://www.datafountain.cn/competitions/350/datasets)中下载数据集, 并解压在./data目录。 ## 数据预处理 ```shell cd data python preprocess.py cd .. ``` ## Bert-base 模型 ```shell bash run_bert.sh #5 fold取平均 python combine.py --model_prefix ./model_bert --out_path ./sub.csv ``` ## Bert Whole Word Masking 模型 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 ```shell bash run_bert_wwm_ext.sh python combine.py --model_prefix ./model_bert_wwm_ext --out_path ./sub.csv ``` ## XLNet-mid 模型 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 ```shell bash run_xlnet.sh python combine.py --model_prefix ./model_xlnet --out_path ./sub.csv ``` ## Roberta-mid 模型 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 ```shell 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 ```