How much is the recommended appropriate size for the image dataset. Additionally, how many epochs should be processed for a properly trained data set?