Can you please guide utilizing Validation Data along with an image data store?

2023 6 5
1 个回答
回答已采纳
更新时间：2023 6 12
6 次查看（30 天）

0 个投票

I'm using a data store of 50 k images of Breast Cancer.
I want to verify if my training learning is susceptible to overfitting.
But I don't understand, How to use validation data.
This my script without validation data:

请先登录，再回答此问题。

 采纳的回答

Amanjit Dulai
Amanjit Dulai 2023-6-5

0 个投票

You can specify validation data in the training options with the ValidationData option. Below I've used your test data set, but you probably might want to create a separate validation set from your original data:
opts = trainingOptions("sgdm", ...
"InitialLearnRate", 0.0001, ...
"MaxEpoch", 10, ...
"Plots", "training-progress", ...
"ValidationData", resizeBreastCancerTest);

4 个评论
显示 2更早的评论 隐藏 2更早的评论

Matthieu
Matthieu 2023-6-5
Thank you for your assistance. I understand that you mentioned creating a new data store. Actually, I create a new datastore with the function splitEachLabel named ValData. However, I am unsure how to choose the number of validation images.
In addition, my script takes a long time to compile with the CPU.
Amanjit Dulai
Amanjit Dulai 2023-6-6
There are no definite rules on how to choose the number of validation images, but if you have thousands of observations in your data, a common ratio is to use 70% for training, 15% for test, and 15% for validation.
If you have validation data, at certain points during the training the accuracy will be measured on the entire validation set. This can take a very long time (because the entire validation set is typically much larger than each mini-batch that is used for training). You can speed things up by reducing the number of evaluations on the validation set. This can be done by increasing the ValidationFrequency option (the default value is 50):
opts = trainingOptions("sgdm", ...
"InitialLearnRate", 0.0001, ...
"MaxEpoch", 10, ...
"Plots", "training-progress", ...
"ValidationData", resizeBreastCancerTest, ...
"ValidationFrequency", 500);
Matthieu
Matthieu 2023-6-9
Thanks for your assistance.
I understand your point of view.
I had another question I tried my script.
However, my training progress is not efficient. I would like to know if I'm in a case of overfitting. Could you please give me some advice in order to improve the results?
Amanjit Dulai
Amanjit Dulai 2023-6-12
There are a few things you could try.
  • The initial learning rate of 0.0001 is quite low. You could try increasing it.
  • Alternatively, you could try using the ADAM algorithm for training, which tends to be less sensitive to hyperparameters. You can do this by changing "sgdm" in the call to trainingOptions to "adam".
  • You could train for more epochs. 10 epochs is a fairly low number.

请先登录，再进行评论。

更多回答（0 个）

类别

在 帮助中心 和 File Exchange 中查找有关 Biotech and Pharmaceutical 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by