search

D641593 / MixNet

MIT License
60 stars 9 forks source link

The model is harder to converge with `mid` mode (--mid True) #4

Open huyhoang17 opened 11 months ago

huyhoang17 commented 11 months ago

@D641593

Thanks for the great repo. I trained 2 model versions: with and without mid mode (--mid True/False). Seem like the model with the midline prediction branch (--mid True) is harder to converge than the other one. I trained the model from scratch on the Totaltext dataset and did not pre-train on Synthtext. Do you have any advice about this issue? Thank you. Hope to see your response soon!

D641593 commented 11 months ago

Hi Please try the following steps

  1. Use --mid False to train on the total-text.
  2. Load the optimal weights from the above steps and train for a few epochs using --mid True.

Applying --mid True should improve the evaluation results.

Can you give me the evaluation result for --mid False? F1-score should be between 86%-88%. If the result is below the range, there's something wrong with my code.