How do I train a CNN using my own dataset?

Question

I'm trying to run an image classification CNN on my own set of labeled images. Currently, I have the images of each label stored in folders, named by the labels. I can also store the images as numpy arrays of pixels. How should I be storing the labeled images so that I can feed them to a CNN?

All the examples I can find online for CNN image classification are using a preloaded dataset from Keras, such as MNIST. I'd like to use my own labeled images, but I can't find any resources on how to do that. Any help would be greatly appreciated!

Answer 1

That largely depends on the framework you're using. For Torch, there's torchvision.datasets.ImageFolder with ToTensor() transformation, for TensorFlow, you'll probably use tf.keras.preprocessing.image.ImageDataGenerator.flow_from_directory(). Both would create a dataset with labels based on subdirectory names, the images may be provided in any common format.