Develop learner multilabel

[bstellato]

Add option to have multiple labels in the PyTorch learner. This is helpful when the strategy for the integer variables corresponds to the actual integer variables vector. (Similar holds for active constraints).

This PR also adds the n_layers as parameter.

TODO

• [x] Implement onehot flag to initialize learner. This also adapts the learner cost function.
• [ ] Fix issue with train/evaluation metric during the training!
• [ ] Implement logic to choose from the problem size whether to use the onehot encoding or the multiple labels for active constraints and integer variables.
  • Split the predictors in one for the active constraints and one for the integer variables
  • Smartly choose between onehot (when there are many variables/constraints and few different combinations from data) and multiple labels (small number of variables/constraints and many different combinations data)

[juliayyan]

So I was running this and got a runtime warning:

RuntimeWarning: overflow encountered in exp
  normalized_outputs = np.reciprocal(1 + np.exp(-outputs))  # Normalize using sigmoid

It's still running so I can't confirm, but I think the issue might be here: https://github.com/bstellato/mlopt/blob/develop_learner_multilabel/mlopt/learners/pytorch/pytorch.py#L92 https://github.com/bstellato/mlopt/blob/develop_learner_multilabel/mlopt/learners/pytorch/pytorch.py#L104

Unlike the one-hot case, I need to wrap the sigmoid around the net outputs to get 0-1s: https://github.com/bstellato/mlopt/blob/develop_learner_multilabel/mlopt/learners/pytorch/pytorch.py#L299 which is not being done before the calls to the MSE function.

[bstellato]

I think it is just a warning. A solution could be this little trick which needs to be adapted to the vectorized case. Another solution is to silence that working inside the function call. In any case it should not affect the output of your code.