search

LinXueyuanStdio / LaTeX_OCR_PRO

:art: 数学公式识别增强版：中英文手写印刷公式、支持初级符号推导（数据结构基于 LaTeX 抽象语法树）Math Formula OCR Pro, supports handwrite, Chinese-mixed formulas and simple symbol reasoning (based on LaTeX AST).
GNU General Public License v3.0
1.07k stars 234 forks source link

configs/training.json doesn't work #62

Open aspnetcs opened 2 years ago

aspnetcs commented 2 years ago

configs/training.json is useless, changing the learning rate in it does not work at all, the learning rate I have been using is CosineAnnealingLR

LinXueyuanStdio commented 2 years ago

do you change lr_init? lr_init is the initial learning rate, while CosineAnnealingLR is learning rate scheduler

aspnetcs commented 2 years ago

yes, LaTeX_OCR_PRO/configs/training.json This file seems to be useless

--

At 2022-02-28 10:40:42, "兮尘" @.***> wrote:

do you change lr_init? lr_init is the initial learning rate, while CosineAnnealingLR is learning rate scheduler

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

LinXueyuanStdio commented 2 years ago

please refer to model/utils/lr_schedule.py, which defines object LRSchedule. Warming-up (lr_warm, end_warm) and decay (start_decay, end_decay) are taken to schedule the learning rate. the learning rate will be lr_init only when the epoch has end_warm < epoch < start_decay

aspnetcs commented 2 years ago

I feel particularly complicated, how to use pytorch's lr_scheduler.MultiplicativeLR for training? thanks！！！

--

At 2022-03-01 15:31:00, "兮尘" @.***> wrote:

please refer to model/utils/lr_schedule.py, which defines object LRSchedule. Warming-up (lr_warm, end_warm) and decay (start_decay, end_decay) are taken to schedule the learning rate. the learning rate will be lr_init only when the epoch has end_warm < epoch < start_decay

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

LinXueyuanStdio commented 2 years ago

you can refer to https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.MultiplicativeLR.html for pytorch's MultiplicativeLR.

LinXueyuanStdio commented 2 years ago

emmm... it seems you are trying to reproduce with pytorch? lr_scheduler is only a trick to improve performance. you can try any LR on your own. Or you can even give up using LR and use a fixed learning rate.