I want to develop a CNN model to identify 24 hand signs in American Sign Language. I created a custom dataset that contains 3000 images for each hand sign i.e. 72000 images in the entire dataset.

For training the model, I would be using 80-20 dataset split (2400 images/hand sign in the training set and 600 images/hand sign in the validation set). My question is:
Should I randomly shuffle the images when creating the dataset? And Why?

PS: Based on my previous experience, it led to validation loss being lower than training loss and validation accuracy more than training accuracy.

  • 1
    $\begingroup$ The answer is: Yes, you should shuffle the dataset, for an explanation check think link. About the higher validation/training loss: is a consistent over several training trials? Also, is it consistent over several epochs or does it happened only at the beginning of each training? $\endgroup$ Apr 13, 2020 at 21:18
  • $\begingroup$ The validation loss was low throughout the training phase. You can check it here= ai.stackexchange.com/questions/18551/… $\endgroup$ Apr 13, 2020 at 21:57
  • 1
    $\begingroup$ @EdoardoGuerriero That looks a lot like an answer rather than a comment. Please consider expanding it to an answer, and adding a little extra detail about what's in the link. $\endgroup$ Apr 14, 2020 at 23:40


You must log in to answer this question.