I am running the same code training the same CNN model using the same dataset on GUP and CPU, and I am using k-fold cross validation in my code. The problem is that k-fold seems not working properly on GPU, because on CPU the number of samples that was used for training after the cross validation is about 700 samples in each fold. However, on GPU it is only 27 samples used for training in each fold.
I don't know what is the problem? could someone please help me with this?