I am looking at the output of a multi-class image segmentation deep learning model. I used U-Net to implement this.

I am confused about why the training accuracies are different for a different validation dataset when using the same training dataset and exactly the SAME model.

Note: The hyperparameters are set as same for both cases. a) Adams optimizer: lr = 0.001 b) batch size = 1 (Yes, I know I will need to use a bigger batch size; this is only a test) c) epoch = 1000

The only difference is a) Augmented uses 48 images for validation b) Pseudo-Augmented use of 2 images for validation.

Thank you in advance.

The output is shown below. enter image description here

  • $\begingroup$ Hi @user10529827, welcome to the site. Are these differences consistent across multiple runs? $\endgroup$
    – noe
    Feb 20 at 16:50
  • $\begingroup$ @noe Yes, they are very similar. Does the validation dataset (size) affect the training accuracies? $\endgroup$ Feb 20 at 17:33
  • 1
    $\begingroup$ No, but the randomness does. I was thinking that maybe the results were not consistent across runs due to different parameter initializations. $\endgroup$
    – noe
    Feb 20 at 17:40
  • $\begingroup$ @noe Thank you for your answer. What do you mean by the randomness? The randomness of the data split? I have a random seeding set to a variable, so the dataset is consistent between the two different sets. $\endgroup$ Feb 20 at 18:04
  • $\begingroup$ There are multiple sources of randomness: the data split, the training data ordering, the weight initialization and other elements (e.g. dropout). $\endgroup$
    – noe
    Feb 20 at 18:38


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.