About 'is_legacy' in config

[Changlin-Lee]

As shown in the training config, 'is_legacy' flag is missing, resulting 'is_legacy' is set to be True in training stage. Meanwhile, in the infer stage, 'is_legacy' is set to False in the config.

Why the training stage and infer stage is not aligned in that case?

[cornettoyu]

thanks for your interest. The is_legacy flag is used to distinguish the two-stage training models (as reported in the paper) and our new released models with an improved single-stage recipe. The training configs are only for two-stage training now and please stay tuned :)