parseq
parseq copied to clipboard
How to reproduce the accuracy of Table17 in the Appendix J?
Hello~ Thanks for you great work. I try to reproduce the parseq in for Table 17 in Appendix J, using MJ and ST with A100 40GB * 2. I got low accuracy about 2.99%. Do you add rotation augment for training?

To reproduce Table 17, you need to use a dataset which contains arbitrarily-oriented text (e.g. the real datasets). Text in MJ+ST are primarily horizontally-oriented.
I guess you could also augment MJ+ST with a fixed rotation by passing +data.rotation=90 (or any non-zero rotation angle) to train.py.
Thanks!That is Cool ~ Thank you again for sharing and making open source available. 👍