Deep Learning with PyTorch

Data representativeness

In the example we saw in our last chapter, we classified images as either dogs or cats. Let's take a scenario where all the images are sorted and the first 60% of images are dogs and the rest are cats. If we split this dataset by choosing the first 80% as the training dataset and the rest as the validation set, then the validation dataset will not be a true representation of the dataset, as it will only contain cat images. So, in these cases, care should be taken that we have a good mix by shuffling the data before splitting or doing a stratified sampling. Stratified sampling refers to picking up data points from each category to create validation and test datasets.