Jianquan Ye

> I can't find the file by following the path ./work_dir/1114_TableMASTER_structure/ 8. Is that file private? It's a custom setting, you should change this path to your work_dir path

> I want to train in my language . What is config i must modify ? `if label_convertor['dict_type'] == 'DICT90': PAD = 92` What is pad = 92 mean ....

> @JiaquanYe Thank you for reply . Can model recognition space ? I want to add space character in DICT90. Sure, you can add space character in your custom dictionary,...

需要自己补充数据训练，这套算法框架做这个没问题 发自我的iPhone ------------------ 原始邮件 ------------------ 发件人: DZ ***@***.***> 发送时间: 2021年12月14日 14:43 收件人: JiaquanYe/TableMASTER-mmocr ***@***.***> 抄送: Subscribed ***@***.***> 主题: Re: [JiaquanYe/TableMASTER-mmocr] 请问对于倾斜的 手机拍照图片可以识别吗？比如下面的图片 (Issue #35) — You are receiving this because...

> 你好，请问你们有测试过其他公开数据集的性能吗？比如ICDAR2013，SciTSR等 这些数据集的标注格式并不是HTML格式，请问有方法能进行转化吗？ Sorry, I have not try this method on others dataset.

I recommend run single process to make sure there is no any running error, before you run multi-process mission.

> 通过解码预测得到的框大概率会有偏移，有什么办法可以优化这种情况吗？ Hi, could you tell me which dataset is used to test? In my experiment, the output box of this model is good when I test pubtabnet dataset.

> @RogerYu123 我用预训练的模型在 pubtabnet 的数据上跑，结果也会有偏移，你这边正常么？  It is abnormal. maybe you can check your visualization code.

1. 原意是一个是代表 td对（里头没有空格），一个代表 eb对（里头没有空格） (可以参照开源代码的structure alphabet，和data_preprocess.py中empty_bbox_token_dict的映射关系 2. 由于colspan 和 rowspan的存在，而把其彼此分开了

I think it is normal if the loss can converge. grad_norm nan is cause by mixed precision training.