Maybe I’m missing something or I use the wrong search phrases, but I can’t find any solution – thanks for your help.

I’m planning to implement an image database (using MongoDB / GridFS) to serve training data for several Image Classification tasks, each using a different subset of images. Labels will be stored in MongoDB as well.

For Classification, I want to use Keras / TensorFlow / tf.data, but my problem is, I find nothing about how to feed the images into tf.data in an efficient / elegant way. There are hints how to store the model (nice), but nothing about how to use the images for Training. It’s always only images on disk, but this seems quite limiting and errorprone for me…

Thank you for any helpful hint or example!