LBH1024 / CAN

When Counting Meets HMER: Counting-Aware Network for Handwritten Mathematical Expression Recognition (ECCV’2022 Poster).
MIT License
364 stars 59 forks source link

训练较大的数据集 #11

Closed Rechargeablezz closed 2 years ago

Rechargeablezz commented 2 years ago

您好,请问如果想在该模型上训练类似HMER100K的数据集,需要调整什么超参数吗

LBH1024 commented 2 years ago

您好,请问如果想在该模型上训练类似HMER100K的数据集,需要调整什么超参数吗

你好,在HME100K上可以将weight_decay调小一些,除了backbone之外的dropout可以设置为false,因为在大数据集上不容易过拟合。

Rechargeablezz commented 2 years ago

好的,非常感激您的建议