Bartzi
commented
5 years ago I'm sorry but I do not speak chinese and I'm definitely not able to read it :sweat_smile: could you ask your question again in English?
Open MS-MA opened 5 years ago
Bartzi
commented
5 years ago I'm sorry but I do not speak chinese and I'm definitely not able to read it :sweat_smile: could you ask your question again in English?
MS-MA
commented
5 years ago oh,i am sorry, i didn't realize this problem.
@Bartzi 您好： 我有一个关于fsns模型训练的问题，恳求得到您的回答。 在准备fsns训练数据时，做完准备数据的1、2、3、4四步后，我理解的是 此时每张图片中单词数最多为默认的6个，每个单词最多有21个字符。但单词中字符的个数不足21个时，gt文件就用blank_label的类(133)填充，单词数不足6个时，gt文件也用类133填充。但是在创建curriculum learning strategy时，将从tfrecord提取的groundtruth based on line的gt文件转为groundtruth based on correspond nummber of words的新的gt文件时， 又用类0代表blank_label，在transform_classes.py文件中用类0作为间隔来区分单词，不足21个字符的单词用类0进行填充，这么操作，我会认为数据准备阶段的3和4两步会起不到任何作用，可以不再去执行。不知道我这样理解的是否正确