Liu Danfeng

你的格式转错了可能，converts里面有转SynthText的文件，直接用就行。

我的tensorflow 1.9.0 keras 2.1.5训练完全没有问题。

我用一张V100训练50K，大概一晚上就够了。这个任务其实对CPU占用比较大。

> @ldfinfontainebleau 您好，我想请教下，您是用自己的数据集做训练的么？ 用的公共数据，CTW+SynthText

It seems value of TEST.DATASETS in mask_textspotter.yaml is not correct.

I write my generator and ignore bg_img, it is training now. Everything seems fine.

> > > what dataset can be used in this repo? Do you mean data of the training program? I used it on hand key points. My training data just...

> > > How do you get the data set? Whether to label yourself or use the official data set? thks It's private data(about 200000+ imgs, company property). You could...

> > > > > what dataset can be used in this repo? > > > > > > Do you mean data of the training program? I used it...

step by step... It has been a while. I will give it a try. 1. make sure your dataset has two parts : .jpg and landmarks 2. I did fix...