I know it's a broad question, sorry for that, but I'm still testing the waters with machine learning.

I have a typical regression task (predict target numbers with the help of features x,y,z) and a dataset with about 1000 entries. I want to use this task to compare several machine learning libraries.

My current selection is Spark, PyTorch, TensorFlow, scikit. Is this selection okay? Do you have further recommendations? My only requirements are that they are open source.

Thanks for your help!

  • $\begingroup$ Microsoft CNTK is also open source. docs.microsoft.com/en-us/cognitive-toolkit $\endgroup$ – serali Nov 17 at 19:24
  • $\begingroup$ Hi, what is it you{re trying to compare? Are you going to run the same model and are looking for different implementations? Are you measuring performance, or accuracy or something else? Regarding your choices so far - AFAIK, spark is not really a library, it's a distributed computing framework, but it does have a library sparkML. Tensorflow and Pytorch are used to develop deep learning networks, with especially Pytorch having multiple features and functions that make it easy to build convolutional networks and such, but I don't see how all that is relevant to a simple regression task.... $\endgroup$ – leon dobrzinsky Nov 18 at 5:52
  • $\begingroup$ Hi @leondobrzinsky, yes I want to compare the implementations for several tasks. I'm doing a project paper on ML, and my main goal with this paper is to get an overview of commonly used open-source libraries. It seems you think these tools are far too overpowered for what I want, and you're probably right - I'm just trying to get a foot in the door. Or do you think that their implementations are too similar for simple regression tasks (think: linear regression, SVM, polynomal regression etc)? Thanks for your time! $\endgroup$ – maRei Nov 18 at 9:17
  • $\begingroup$ Well. kind of both. I'm far from an expert on the subject, but pytorch and tensorflow both deal with nerual networks, so the differences for a linear regression task would have more to do with the parameters you provide and the model you build and not with their implementation. Also, they both are kind of overpowered and not best fit for this task. $\endgroup$ – leon dobrzinsky Nov 22 at 17:13
  • $\begingroup$ Furthermore, take into account that most libraries provide quite a few different ways t build and implement a model, so part of the differences will be due to your choices, or to the default settings. If you really want to compare exactly the same model in different libraries, I would suggest looking for libraries built with different programming languages maybe...Again, I think you need to understand exactly which parameters you want to change and compare. Sorry I can't provide better guidance, but at least I hope it gives foodfor thought... $\endgroup$ – leon dobrzinsky Nov 22 at 17:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.