Iteration setting in the paper and the code

[dalindalin]

hey, I have some questions about the number of training iterations.
The paper says when training on ICDAR2015, you first pre-train on Synthtext for 60k iterations, then train on ICDAR2015 for 8k iterations in the stage 1 and 4k iterations in the stage 2. then

1. What's the batch_size for each iteration in the paer? is iteration == epoch in your paper?
2. the number of iterations is diffrent from the iteration parameters in the modelConfig.py, which is 40k/100k-40k/120k-100k with a batch_size of 32, so which experiment is the default iteration setting for? Can I reproduce the result on ICDAR2015 in the paper with the same setting?

[MhLiao]

1. The batch size is set to 32. 1 iteration means 1 batch, not an epoch.
2. You can refer to the paper to adjust your settings.

[dalindalin]

got it,thank you~ @MhLiao

[basaltzhang]

so the batch size of 384x384 and 768x768 are both 32, is that right?

[ZDDEAN]

嘿，我对训练迭代次数有一些疑问。
该论文称，在ICDAR2015上进行培训时，首先在Synthtext上进行6k次迭代预训练，然后在第2阶段训练ICDAR2015进行8k次迭代，然后在第2阶段进行4k次迭代训练。

1. paer中每次迭代的batch_size是什么？是你的论文中的迭代==时代？
2. 迭代次数与modelConfig.py中的迭代参数不同，后者为40k / 100k-40k / 120k-100k，batch_size为32，所以哪个实验是默认的迭代设置？我可以使用相同的设置在ICDAR2015中重现结果吗？

请问如何能分享一下synthtext格式转换为icdar格式的脚本吗，谢谢鸭