PST-table
PST-table copied to clipboard
表格结构解析新思路(表格识别新思路)
can you send it by email? mail: [email protected] please!!!
能提供份数据么?
文件那个百度云的链接能不能提供下,好像被删除了
why the loss use log_pointer_score not argmax_pointer,but train_accuracy use argmax_pointer unrolled = log_pointer_score.view(-1, log_pointer_score.size(-1)) loss = F.nll_loss(unrolled, target.view(-1), ignore_index=-1)
没有 dataset_refined_ocr_test 这个文件
请问代码中是对 down_to_up, up_to_down,right_to_left,left_to_right四个方向的pointer单独训练的吗?因为看到dataset_refined_ocr里面parent只取了一个值
Could you give some examples for tab_post.py? ```python ocred_path = '/home/gita/Downloads/mini_result/mini_json/' img_path_txt = '/home/gita/Downloads/mini_result_50/mini_father.txt' uf_path = '/home/gita/Downloads/mini_result_50/father/' df_path = '' lm_path = '/home/gita/Downloads/mini_result_50/mother_p/' rm_path = '/home/gita/Downloads/mini_result_50/mother_n/' ``` I don't kown...
tab_pre.py代码中表述的可能是合并单元格后单元格内部换行,参考:.\pubtabnet\train\PMC1626454_002_00.png  通过横向投影直方图确定有几个H_Start,如果不为1才要进行后续处理,所以可能是这个思路
请问./mode下文件有?是什么文件?
请问这个ocr识别是怎么样的?