[FEA] Confusion matrix

[beckernick]

It would be great to be able to calculate the confusion matrix of the predictions/ground truth of a given estimator. Calculating the confusion matrix from preds/ground truth of size ~3 million can take at least 2.5 seconds on the CPU with sklearn.

Following mention in the text of #242 and #608 , filing an issue to make tracking explicit.

[beckernick]

Related to #242 , the following would be a basic 2-class confusion matrix with CuPy:

def cupy_conf_mat(y, y_pred):
    """
    Simple, fast confusion matrix for two class models designed to match sklearn.
    Assumes the classes are one of [0, 1]. It will fail edge cases, which are fairly
    numerous.
    
    Could be expanded to multi-class by following similar logic to the precision implementation,
    but no need for now.
    
    This is about 300x faster than sklearn for 3M float64 predictions.
    """
    nclasses = len(cp.unique(y))
    assert nclasses == 2
    res = cp.zeros((2, 2))
    
    pos_pred_ix = cp.where(y_pred == 1)
    neg_pred_ix = cp.where(y_pred != 1)
    tn_sum = (y[neg_pred_ix] == 0).sum()
    fn_sum = (y[neg_pred_ix] == 1).sum()
    tp_sum = (y[pos_pred_ix] == 1).sum()
    fp_sum = (y[pos_pred_ix] == 0).sum()
    
    res[0, 0] = tn_sum
    res[1, 0] = fn_sum
    res[0, 1] = fp_sum
    res[1, 1] = tp_sum
    return res