/* 

  Train the Spam data set in Picat.

  Right now this is only for categorical datasets...

  For spambase, all predictions are - what I can see - the same:
    {0.582182288169861,0.418089509010315}]

  Some other tests:

  - NN = new_sparse_nn({NumInputs,50,NumOutputs}, 0.34)
    it's a little better
    result = [num_correct = 429,num_tests = 460,pct = 0.932608695652174]

    Confusion matrix:
     {278,11}
     {20,151}

  - NN = new_sparse_nn({NumInputs,50,NumOutputs}, 0.73)
    result = [num_correct = 418,num_tests = 460,pct = 0.908695652173913]

    Confusion matrix:
    {262,27}
    {15,156}

  - NN = new_sparse_nn({NumInputs,50,2, NumOutputs}, 0.35),   
    nn_set_activation_function_hidden(NN, linear),

    result = [num_correct = 217,num_tests = 230,pct = 0.943478260869565]

    Confusion matrix:
     {141,5}
     {8,76}


  This Picat model was created by Hakan Kjellerstrand, hakank@gmail.com
  See also my Picat page: http://www.hakank.org/picat/

*/

import nn.
import nn_hakank.

main => go.

go =>
  Base = "spam.arff",
  Map = new_map([1='nospam',2='spam']),
  train_and_test(Base, Map),
  nl.


%%
%% train_net(TrainFile, NetFile,NumInputs,Numoutput)
%%
%% This must be defined in order to train_and_test/2 to work.
%% train_net/4 is called from train_and_test/2.
%%
train_net(TrainFile, NetFile,NumInputs,NumOutputs) =>

  % {num_inputs, num_hidden_layers..., num_outputs}
  %%NN = new_nn({NumInputs,40,NumOutputs}),
  NN = new_sparse_nn({NumInputs,50,2, NumOutputs}, 0.35),   
  nn_set_activation_function_hidden(NN, linear),
  nn_train(NN,TrainFile,$[maxep(5_000), report(1000), derror(0.0), bfl(0.1), train_func(rprop), stop_func(stop_bit)]),
  nn_save(NN,NetFile),
  % nn_print(NN),
  nn_destroy_all.