识别v3确实会比v2好吗？

[shuizhonghaitong]

你好，我采用相同的数据集分别训练识别的v2和v3。 1、在训练的过程中，我发现在相同的epoch下，v3的效果没有v2好，无论是train acc还是dev acc，这是正常的嘛？v3的效果要超过v2的话是不是v3需要用的epoch要比v2多？v3的学习率是否需要根据数据集的大小做调整？ 2、如果几十万的训练数据，对于v2和v3，各自大约需要跑多少个epoch呢？

[bely66]

I didn't run this test, but I think It's an interesting point of view In their paper they compare the two: https://arxiv.org/pdf/2206.03001

Part of the improvement was in the dataset collection and labelling + augmentation so that might be it

[tink2123]

论文中有给出相同数据下的消融实验，在26w真实训练数据上，v3训练精度明显高于v2，在不使用模型蒸馏的情况下就已经可以超越v2蒸馏模型的精度。

上述问题可能与训练数据有关，场景是否比较单一呢。如果训练场景相对简单，建议训练过程去掉GTC策略，避免过拟合。

几十万的训练数据，跑200左右epoch即可。

[shuizhonghaitong]

场景是文档数据识别，我去掉了GTC，但是效果仍然没有超过v2

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.