Why am I getting different accuracy in training a deep CNN in each trial?

Hi there

I am using deep CNN architecture for classsification task. I wonder why I am getting different accuracy every time. Is there any way to find the best accuracy out of this model? As I cant run it over 100 times for such long epochs on cpu :(

or is the the only way to calculate accracy is to find mean of all these trials

plz guide.

Thanks

Answer 1

[Ameer Hamza]

The initialization of the deep learning model is random. It is expected that every time you run the training, the result will be somewhat different. But it should not be significantly different. If they are, then this shows a problem with your model, which is too sensitive to the initialization. If the difference is not that significant, then you can just take the best value.

[krishna Chauhan]

@Ameer Hamza,Thank you.

Oh then How to remove this model issue?

actually its quite varrying

i just complted three runs

53.36

52.44

51.15

so what to do now?

[Ameer Hamza]

These look close enough. Choose the best one.

By issues with the model, I meant issues such as underfitting.

[krishna Chauhan]

okay are you finding its okay?

no the model seems to be overfitting ..as I can see from training graph here.

thats why m using early stopping as of now

[Ameer Hamza]

If there is overfitting then try using a smaller model. Also, what do these value represent? Are these accuracy values?

[krishna Chauhan]

Smaller model meaning, batch size needs to be reduced?

yes this is accuracy only.

[Ameer Hamza]

By smaller model, I mean a model with fewer layers. Or less number of neurons per layer.

Accuracy of around 50% does not seem good. It is equivalent to a coin-toss. The network is essentially giving random values.

[krishna Chauhan]

This datset is quite complex , state of art is between 65-70 %for my 5 layer network its quite good. If random values then it cud take anything why betweeen 50 and 60. Plz guide

[Ameer Hamza]

I do agree that it is slightly above 50% in all the cases, so it is not the same a random value. But it still seems low. It depends on your requirement and state of the art. If the accuracy of 53% makes sense in your case, then it should be fine.