Replies: 2 comments
-
根据您提供的训练日志对比,我建议不要在训练初期就加入疑难近似字符。原因如下:
这样的渐进式训练策略可以帮助模型更好地学习处理疑难近似字符,同时保持对基础字符的良好识别能力。 Response generated by feifei-bot | claude-3-5-sonnet-latest |
Beta Was this translation helpful? Give feedback.
0 replies
-
如果在预训练的svtr2模型上改变字典,增加到几万个字符,恢复训练,
与全新训练几万个字符,哪种效果会好?
…--------------------------------------------------------------------------------
------------------ 原始邮件 ------------------
发件人: Wang Xin ***@***.***>
发送时间: 2025-01-10 09:40:01
收件人:PaddlePaddle/PaddleOCR ***@***.***>
抄送:nissanjp ***@***.***>,Author ***@***.***>
主题: Re: [PaddlePaddle/PaddleOCR] svtr2 疑难近似字符应该在什么时候加入训练? (Discussion #14523)
根据您提供的训练日志对比,我建议不要在训练初期就加入疑难近似字符。原因如下:
从训练效果来看: 不加入疑难近似字符时:acc达到0.5641,norm_edit_dis达到0.8391 加入15%疑难近似字符后:acc下降到0.3422,norm_edit_dis下降到0.6666 建议的加入时机: 建议在模型基本收敛,识别准确率相对稳定后再逐步加入疑难近似字符 可以采用渐进式方式:先用基础字符训练至一定准确率→再少量引入疑难字符(如5%)→观察效果后再逐步增加比例 避免一次性加入过多(如15%)疑难字符,这样可能会导致模型难以收敛 具体建议: 先用基础字符训练直到acc稳定在0.6以上 每次只增加约5%的疑难近似字符 每次加入新字符后,给予充分的训练轮次让模型适应 观察验证集指标变化,如果指标下降明显则适当调整策略 这样的渐进式训练策略可以帮助模型更好地学习处理疑难近似字符,同时保持对基础字符的良好识别能力。
Response generated by feifei-bot | claude-3-5-sonnet-latest
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
svtr2 疑难近似字符应该在什么时候加入训练?
加入疑难近似字符前日志如下
[2025/01/10 08:46:42] ppocr INFO: epoch: [1/50], global_step: 800, lr: 0.000001, CTCLoss: 14.587288, NRTRLoss: 2.853578, loss: 18.047794, avg_reader_cost: 0.00208 s, avg_batch_cost: 0.18747 s, avg_samples: 2.35, ips: 12.53550 samples/s, eta: 26 days, 8:36:20, max_mem_reserved: 2097 MB, max_mem_allocated: 1821 MB
[2025/01/10 08:47:01] ppocr INFO: epoch: [1/50], global_step: 900, lr: 0.000001, CTCLoss: 28.381334, NRTRLoss: 3.626843, loss: 32.118370, avg_reader_cost: 0.00200 s, avg_batch_cost: 0.17118 s, avg_samples: 2.22, ips: 12.96876 samples/s, eta: 25 days, 23:04:52, max_mem_reserved: 2097 MB, max_mem_allocated: 1821 MB
[2025/01/10 08:47:19] ppocr INFO: epoch: [1/50], global_step: 1000, lr: 0.000001, CTCLoss: 18.424065, NRTRLoss: 2.977797, loss: 21.543367, avg_reader_cost: 0.00181 s, avg_batch_cost: 0.17177 s, avg_samples: 2.41, ips: 14.03078 samples/s, eta: 25 days, 15:38:51, max_mem_reserved: 2097 MB, max_mem_allocated: 1821 MB
[2025/01/10 08:48:30] ppocr INFO: cur metric, acc: 0.5641729542609956, norm_edit_dis: 0.8391193097812136, fps: 23.73929370291974
加入15%疑难近似字符后日志如下
[2025/01/10 07:08:17] ppocr INFO: epoch: [1/50], global_step: 800, lr: 0.000001, CTCLoss: 27.405363, NRTRLoss: 3.751744, loss: 31.468330, avg_reader_cost: 0.00612 s, avg_batch_cost: 0.18858 s, avg_samples: 2.42, ips: 12.83303 samples/s, eta: 24 days, 15:34:07, max_mem_reserved: 2086 MB, max_mem_allocated: 1817 MB
[2025/01/10 07:08:35] ppocr INFO: epoch: [1/50], global_step: 900, lr: 0.000001, CTCLoss: 34.598011, NRTRLoss: 4.159418, loss: 39.047333, avg_reader_cost: 0.00216 s, avg_batch_cost: 0.17124 s, avg_samples: 2.32, ips: 13.54832 samples/s, eta: 24 days, 10:37:32, max_mem_reserved: 2086 MB, max_mem_allocated: 1821 MB
[2025/01/10 07:08:54] ppocr INFO: epoch: [1/50], global_step: 1000, lr: 0.000001, CTCLoss: 25.757986, NRTRLoss: 3.991220, loss: 29.173361, avg_reader_cost: 0.00482 s, avg_batch_cost: 0.18058 s, avg_samples: 2.27, ips: 12.57071 samples/s, eta: 24 days, 9:39:15, max_mem_reserved: 2086 MB, max_mem_allocated: 1821 MB
[2025/01/10 07:10:51] ppocr INFO: cur metric, acc: 0.3422069543056807, norm_edit_dis: 0.6666195348935893, fps: 23.206885368755316
Beta Was this translation helpful? Give feedback.
All reactions