PyTorch is research favorite, it's easy to debug tensors and interleave bunch of random code in between layers if you need to. Keras is the easiest one to get a

PyTorch is research favorite, it's easy to debug tensors and interleave bunch of random code in between layers if you need to. Keras is the easiest one to get a model running in production, but for some advanced things like sophisticated custom losses you might be forced to switch over to base TensorFlow. TensorFlow is the production king but it's super complex unless you use tf.keras and difficult to debug.