I get MLP implementation from scikit-learn and compared it with TensorFlow implementation. I used the same hyperparameters and the same optimizer (Adam). I've compared implementations on 114 datasets from Penn ML Benchmarks - they represent binary classification, multiclass classification, and regression. In comparison, I was using the CPU only. The TF was only 59 times out of 114 better than sklearn. I was surprised. The differences in performance weren't huge. What is more, the sklearn implementation seems to be about 2 times faster (I was using CPU only).<p>The conclusion is, that if you are already using Sckit-learn and need classic MLP then there is no need to get TensorFlow.