I have been using 'nnet' in 'caret' to train neural networks. 'nnet' allows only one hidden layer and only sigmoid activation function, which restricts from exploring different activation functions and no. of hidden layers.
I generally train neural nets for classification problems on tabular datasets with 1,000 to 10,000 rows and 10-25 columns. I can see there are many neural network packages - mxnet, neuralnet, deepnet, keras, tensorflow. h2o, etc. I cannot try all of them and have to decide and learn one.
- Can someone recommend me on which library should I use.
- Do these libraries have their own ways to search hyper-parameter, cross-validation or I need to use caret.