Actually, resources are the least of your concerns when it comes to DL. DLs have a shitload of hyper-parameters (effectively, the entire architecture) which is

Actually, resources are the least of your concerns when it comes to DL. DLs have a shitload of hyper-parameters (effectively, the entire architecture) which is mostly built by trial and error + intuition. There was some really exciting work on reversible SGD by an applied math group at Harvard (https://github.com/HIPS/hypergrad https://github.com/HIPS/hypergrad) to obtain derivatives with respect to network architecture, but it's not very mature yet. Further, DLs need an immense amount of data to be efficient. If your dataset has only a few hundred of features, and you have a few thousand samples, it's unlikely DL will help you much. Finally, DL is not really useful out of the box, at all. scikit-learn's selling point is that it has an amazingly good uniform API, fantastic documentation, and very clean code. Making a working DL implementation with a simple 'fit' sklearn-like API is impossible.