inference for augmentedImageDatastore with 'OutputSiz​eMode','ra​ndcrop'

Hi,

Training and evaluating the CNN model has no issue for me if the image input size is the same or bit bigger which can be resize to be same with model input.

My problem is when training with the augmentedImageDatastore options of 'OutputSizeMode','randcrop' where it takes random crop(224x224x3) from the training images which apparently a lot bigger(say 4000x4000x3) then the model input size, i believe the the CNN model now become trained to classify cropped or patches of the original big images and not the class for the whole image.

So how do you correctly infer the whole image with the randcrop options?

thank you.