Mingkun Yang

I guess there should be two reasons, 1. I am not very sure that the 20ms in our paper is whether tested with beam width 1. 2. The max_len in...

> > I guess there should be two reasons, > > > > 1. I am not very sure that the 20ms in our paper is whether tested with beam...

Duplicate of #18

Is the right checkpoint loaded? And do you use the given code for testing or rewrite a new one?

生成数据的代码：`tools/create_svtp_lmdb.py`，格式看代码就懂了。

> I cannot extract the pretrained model and the code shows the same error. You can directly load this pretrained model without extract it.

How about your training dataset and the number of training epochs.

中文的话，直接用这个repo可能效果不是很好。因为中文文字条的长度变化更加大，直接resize到同一个size进行训练的话，可能需要很大的数据量。所以在实际应用中，我一般都是不同batch采用不同长度来训练中文。另外，lr 1看起来好像有点大，可能导致不太好收敛。

At present, the bidirectional attention decoder proposed in ASTER is not included in my implementation.

if possible, you can submit a PR to add the other decoder.