train.py 文件源码

def eval_metrics_on(predictions, labels, regression = True):
    '''
    assuming this is a regression task; labels are continuous-valued floats

    returns most regression-related scores for the given predictions/targets as a dictionary:

        r2, mean_abs_error, mse, rmse, median_absolute_error, explained_variance_score
    '''
    if len(labels[0])==2: #labels is list of data/labels pairs
        labels = np.concatenate([l[1] for l in labels])
    if regression:
        predictions = predictions[:,0]
        r2                       = metrics.r2_score(labels, predictions)
        mean_abs_error           = np.abs(predictions - labels).mean()
        mse                      = ((predictions - labels)**2).mean()
        rmse                     = np.sqrt(mse)
        median_absolute_error    = metrics.median_absolute_error(labels, predictions) # robust to outliers
        explained_variance_score = metrics.explained_variance_score(labels, predictions) # best score = 1, lower is worse
        return {'r2':r2, 'mean_abs_error':mean_abs_error, 'mse':mse, 'rmse':rmse, 
                'median_absolute_error':median_absolute_error, 'explained_variance_score':explained_variance_score, 'main_metric':rmse}
    else:
        predictions = predictions[:,1]


        if labels.max()==1:
            auc    = metrics.auc(predictions, labels[:,1], reorder=1)
        accuracy = np.mean((predictions>0.5)==labels[:,1])

        return {'auc':auc, 'accuracy':accuracy, 'main_metric':accuracy}