FOTS.PyTorch icon indicating copy to clipboard operation
FOTS.PyTorch copied to clipboard

没有进行finutne,训练了1000个epoch,验证:pre:0.05 recall:0.02 hmean:0.03 AP:0?可能是什么问题呢

Open neverstoplearn opened this issue 2 years ago • 6 comments

neverstoplearn avatar May 16 '22 06:05 neverstoplearn

1000 epoch , e2e 不finetune 训练可能有点难。

neverstoplearn @.***> 于2022年5月16日周一 14:41写道:

— Reply to this email directly, view it on GitHub https://github.com/jiangxiluning/FOTS.PyTorch/issues/98, or unsubscribe https://github.com/notifications/unsubscribe-auth/AANWXLSGJBMPM77HAG64Q4DVKHUYHANCNFSM5WANQXCQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

jiangxiluning avatar May 16 '22 07:05 jiangxiluning

1000 epoch , e2e 不finetune 训练可能有点难。 neverstoplearn @.> 于2022年5月16日周一 14:41写道: — Reply to this email directly, view it on GitHub <#98>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AANWXLSGJBMPM77HAG64Q4DVKHUYHANCNFSM5WANQXCQ . You are receiving this because you are subscribed to this thread.Message ID: @.> 预训练模型训练需要时间更久 尴尬

neverstoplearn avatar May 16 '22 07:05 neverstoplearn

1000 epoch , e2e 不finetune 训练可能有点难。 neverstoplearn @.> 于2022年5月16日周一 14:41写道: — Reply to this email directly, view it on GitHub <#98>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AANWXLSGJBMPM77HAG64Q4DVKHUYHANCNFSM5WANQXCQ . You are receiving this because you are subscribed to this thread.Message ID: @.>

不fintune的话大概多少epoch可以有不错的结果啊,我看你在readme里面不finetune 1000轮效果还可以。

neverstoplearn avatar May 16 '22 07:05 neverstoplearn

我最近 训的一个 1000 不finetune 就很拉垮。 Calculated!{"precision": 0.10656316160903317, "recall": 0.07270101107366393, "hmean": 0.0864338866628506, "AP": 0}

jiangxiluning avatar May 16 '22 13:05 jiangxiluning

我最近 训的一个 1000 不finetune 就很拉垮。 Calculated!{"precision": 0.10656316160903317, "recall": 0.07270101107366393, "hmean": 0.0864338866628506, "AP": 0}

结果差不多 啥时候更新一波在SynthText800k预训练模型啊?另外想请教一下,这类端到端的算法是不是不适用于中文OCR?

neverstoplearn avatar May 17 '22 09:05 neverstoplearn

我跑的tf复现的FOTS,效果也不尽人意,检测分支训练效果一直不好。

laofeiwei avatar May 19 '22 03:05 laofeiwei