I am getting different accuracy after each run in DNN.

Hello there

I am using a simple DNN architecture , with 24 layers in total, containing CNN and classification layer only. for 10 epochs I am getting a diffrent accuracy every time is it possible. Also the traing graph is not settled till end of training, I tried for 50 epochs too.

any help?

[Walter Roberson]

I think it likely that there is random initialization going on.

Try using rng() with a fixed seed, such as

rng(655321)

just before each training session, to see whether you get the same result each time.

[krishna Chauhan]

No sir Its not working.

[Kanika Gupta]

Refer to this : https://www.mathworks.com/matlabcentral/answers/24056-every-time-the-neural-network-classification-gives-different-accuracy#answer_31622

[krishna Chauhan]

Sorry I am not sure I get it. Its a simple DNN with only few layers. I m not designing for any complicated network. I have used this rng() before trainNetwork, and ran my model 3 times every time i got different accuracy. How to specify thses hidden nodes?? I m black here.

all m getting is i need the weights for which the accuracy is max, but still ii cant initialize them as i want. so whats gain in that too.

[Walter Roberson]

Can you attach your code and data? You can zip it all up and attach the zip.

[krishna Chauhan]

okay sir I will try. Meanwhile is there any way to run the code over google colab or some other plateform as I get an optimum answer that if I go for say 1000 epochs may be the result will be stable. Currently i was exceuting the code for 10epochs only.

[krishna Chauhan]

I cant attach the whole data its above 5MB. The example I followed was this

https://in.mathworks.com/help/wavelet/examples/spoken-digit-recognition-with-wavelet-scattering-and-deep-learning.html

second part using CNN.

[krishna Chauhan]

Sir kindly find the attachement.I have not included the data folder.

Answer 1

[Aditya Patil]

Set a random seed using the rng function. Otherwise, during each run, a different random seed might be selected, resulting in different results.

[krishna Chauhan]

I did that too before trainNetwork()

But still the problem is same