Does ValidationData in trainingOptions influence training progress

Question

Valentin Steininger 2019-9-27

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/482496-does-validationdata-in-trainingoptions-influence-training-progress

评论： Valentin Steininger 2019-10-1

Hi,

I trained a convolutional network and used ValidationData in trainingOptions to cross-check the validation error during training.

There seems to be a threshold when the RMSE of my training data again increases but still the validation RMSE goes down.

So I was wondering if that option would have an influence on that as there is nothing mentioned in the matlab documentation.

thanks for any help!

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Gaurav Garg 2019-9-30

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/482496-does-validationdata-in-trainingoptions-influence-training-progress#answer_394084

Hi,

The ValidationData in trainingOptions does not have any influence in training progress.

In actual, validation set is used to evaluate the given model and used to fine tune the model hyper-parameters. So, the validation set affects the model, but indirectly.

Also, validation of the data has no effect on training. While, the vice-versa may not be true because the model sees and learns from the training data and the validation model then evaluates the validation data. So, RMSEs of the training and validation phases have no effect on each other.

However, you can use k-fold cross validation to tune/fine-tune your parameters and get the best results for validation and test data. You can use cvpartition function for this.