第一次接触pytorch和bert模型，有几个问题需要请教一下，程序已经跑起来了，但是有几个问题

[shaneNo1]

修改了三处地方：

1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522

2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来

3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')

最终结果：

Train loss = 2.019
Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747
Best Logic Form: 0.732 at epoch 29
Best Execution: 0.747 at epoch 29
Best Mean: 0.740 at epoch 29
Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105)
On epoch individually (28, 28, 29, 20, 24, 28, 29)

好像没有预期的高啊

[vdogmcgee]

修改了三处地方：

1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522

2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来

3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')

最终结果：

Train loss = 2.019
Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747
Best Logic Form: 0.732 at epoch 29
Best Execution: 0.747 at epoch 29
Best Mean: 0.740 at epoch 29
Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105)
On epoch individually (28, 28, 29, 20, 24, 28, 29)

好像没有预期的高啊

谢谢老哥分享 , 第一处修改有点不懂 , 用的是哈工大讯飞的词表 , 跟bert一样是21128 , 为啥要改成30533呢?

[IAMZn1018]

修改了三处地方：
1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522
2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来
3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')
最终结果：
Train loss = 2.019
Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747
Best Logic Form: 0.732 at epoch 29
Best Execution: 0.747 at epoch 29
Best Mean: 0.740 at epoch 29
Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105)
On epoch individually (28, 28, 29, 20, 24, 28, 29)
好像没有预期的高啊

谢谢老哥分享 , 第一处修改有点不懂 , 用的是哈工大讯飞的词表 , 跟bert一样是21128 , 为啥要改成30533呢?

30533是bert英文版本词表大小

[IAMZn1018]

修改了三处地方：

1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522

2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来

3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')

最终结果：

Train loss = 2.019 Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747 Best Logic Form: 0.732 at epoch 29 Best Execution: 0.747 at epoch 29 Best Mean: 0.740 at epoch 29 Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105) On epoch individually (28, 28, 29, 20, 24, 28, 29)

好像没有预期的高啊

第二点在133行 # do mask下面加

[hanyonggihub]

修改了三处地方：

1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522

2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来

3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')

最终结果：

Train loss = 2.019 Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747 Best Logic Form: 0.732 at epoch 29 Best Execution: 0.747 at epoch 29 Best Mean: 0.740 at epoch 29 Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105) On epoch individually (28, 28, 29, 20, 24, 28, 29)

好像没有预期的高啊

修改了三处地方：
1、model\chinese-bert_chinese_wwm_pytorch\config.json, 其中vocab_size的值改为 30522
2、code\sqlnet\model\sqlbert.py,大约141行附近，增加三行：sel_col_mask = sel_col_mask - 254；where_col_mask = where_col_mask - 254；qcol_mask = qcol_mask - 254，程序才能跑起来
3、log日志的位置。code\sqlnet.utils.py文件，950行附近，save_error_case(error_case, gt_cases, dir='./log/')改为save_error_case(error_case, gt_cases, dir='../log/')
最终结果：
Train loss = 2.019
Dev Logic Form Accuracy: 0.732, Execution Accuracy: 0.747
Best Logic Form: 0.732 at epoch 29
Best Execution: 0.747 at epoch 29
Best Mean: 0.740 at epoch 29
Best val acc: (0.9993093922651933, 0.9882596685082873, 0.9861878453038674, 0.975828729281768, 0.9509668508287292, 0.9544198895027625, 0.8031767955801105)
On epoch individually (28, 28, 29, 20, 24, 28, 29)
好像没有预期的高啊

谢谢老哥分享 , 第一处修改有点不懂 , 用的是哈工大讯飞的词表 , 跟bert一样是21128 , 为啥要改成30533呢?

请问一下这个文件怎么获取？
../model/chinese-bert_chinese_wwm_pytorch/pytorch_model.bin

[vdogmcgee]

https://github.com/ymcui/Chinese-BERT-wwm