[英]Should I keep training my model for more Epochs to get a better R Squared value?
This is my model, using Adam optimiser and lr of 0.0001.这是我的 model,使用 Adam 优化器和 0.0001 的 lr。
This is the Training and the Validation loss plot below:这是下面的训练和验证损失 plot:
The value of mse=0.049 and r2=70.3%. mse=0.049,r2=70.3%。 My problem is the more images i include into my training model, the worse it gets but mse and r2 increases so this is very counterintuitive.
我的问题是我在训练 model 中包含的图像越多,它变得越糟糕,但是 mse 和 r2 会增加,所以这是非常违反直觉的。 This is a regression model using Nvidia's Model.
这是使用 Nvidia 的 Model 的回归 model。 I input the images and need a steering angle output.
我输入图像并需要一个转向角 output。 Am i actually making things more difficuilt than it needs to be by increasing data and epochs?
通过增加数据和时代,我是否真的让事情变得比需要的更困难? I am training with only 1600 Images for 30 epochs.
我只用 1600 张图像训练 30 个 epoch。 Will appreciate any suggestion from the experts as im relatively new to NNs.Will provide more details if needed thanks:)
将感谢专家的任何建议,因为我对 NN 来说相对较新。如果需要,将提供更多详细信息,谢谢:)
Observing the graph that shows the results of the training and validation loss we can see that after 12 epochs
(more or less), the validation value does not improve too much as the epochs go up.观察显示训练和验证损失结果的图表,我们可以看到在
12 epochs
(或多或少)之后,验证值不会随着 epoch go 的增加而提高太多。 I think that the most logical thing would be to look for the value from which the validation/training loss does not have much improvement left.我认为最合乎逻辑的事情是寻找验证/训练损失没有太大改善的值。
Also remember that it is important not to overfit the model to the training data, as this can lead to errors when validating.还请记住,不要将 model 过度拟合到训练数据中,这很重要,因为这可能会在验证时导致错误。 I hope my answer will help you.
希望我的回答对你有所帮助。 Good luck!
祝你好运!
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.