hello,

usually, i work with genetic algorithms that have many parameters. At each execution, i find a value of accuracy. Therefore, I pass by setting parameters step.

If the CNN has parameters? If yes, you try to deduce the best. If no, researches use this method: they execute the algorithm 10 times and they compute the mean to find the final accuracy.

Good luck.

Dear Aoautef Mahani ,

did you mean that I try to take the best one?

Hi Toqa,

This happens for several reasons.

Most commonly, the training data size is small, and the data are diverse. As a result, your validation data cannot be considered to have the same distribution as training data. Therefore, your model easily converges to different points and your results vibrate.

Two simple methods can be used to stable your training results. 1) use a data augmentation method to increase the size of training data. But before that, you have to make sure the validation data and training data are from the same data source. 2) pre-train your model uses a large-scale dataset, for example the ImageNet. Then, fine-tune your pre-trained model on your medical dataset. The pre-train strategy can be considered as setting a good starting point.

Best,

Nan

Dear Mr. Nan Meng

I will work on your advice.

Thank you

Hello.

No, you execute your algorithm 10 times and you compute the mean accuracy. This value represent the final result.

Hello.

There is also another efficient method which is called cross validation. In this method, the dataset is divided into k partitions (or folds). for each partition, you use the remaining partitions for training and the actual partition for test. In total, you have k values for accuracy and you take the means. Generally, k=5 is the best value.

Dear Aoautef Mahani ,

Thank you a lot for you answer

All the best,

Toqa