[glimcat]

I routinely do machine learning and computer vision, and have used a number of languages for them.

The main issue, irrespective of language, is that the implementation of your final model is often a distinct step from the last output of your experimental tools. If you just take what it gives you and try to deploy that, it will often provide extremely suboptimal performance.

The methods for implementing your final model could involve raw Python, NumPy, the Python wrapper for an external library, writing and consuming custom C libs...it depends on the complexity of the hyperplanes.

But e.g. scikit-learn already wraps libsvm and liblinear. If your SVM (etc.) is slow, it's very unlikely to be because you used Python.

If you're e.g. trying to do Facebook-level heavy lifting, your experiences may vary. But again, that would be a challenge for any tools. The solution is to use sampling, parallelism, etc. - and to implement and optimize your final model as a separate step from designing it.