LassoRegression.py 文件源码

def evaluate(self):
        """"""
        ## not truncate outliers
        pred_valid = pd.DataFrame(index=self.ValidData.index)
        pred_valid['parcelid'] = self.ValidData['parcelid']

        truth_valid = pd.DataFrame(index=self.ValidData.index)
        truth_valid['parcelid'] = self.ValidData['parcelid']

        start = time.time()

        for d in self._l_valid_predict_columns:
            l_valid_columns = ['%s%s' % (c, d) if (c in ['lastgap', 'monthyear', 'buildingage']) else c for c in
                               self._l_train_columns]

            extra_va = pd.read_hdf(path_or_buf='%s/p21/eval_valid_%s.hdf' % (self.InputDir, d), key='valid')
            #ValidData = self.ValidData.join(extra_va, on= 'parcelid', how= 'left')
            ValidData = pd.concat([self.ValidData, extra_va.drop('parcelid', axis= 1)], axis= 1)

            x_valid = ValidData[l_valid_columns]
            x_valid = x_valid.values.astype(np.float32, copy=False)
            pred_valid[d] = self._model.predict(x_valid)  # * 0.99 + 0.011 * 0.01
            df_tmp = ValidData[ValidData['transactiondate'].dt.month == int(d[-2:])]
            truth_valid.loc[df_tmp.index, d] = df_tmp['logerror']

        score = 0.0
        ae = np.abs(pred_valid - truth_valid)
        for col in ae.columns:
            score += np.sum(ae[col])
        score /= len(pred_valid)  ##!! divided by number of instances, not the number of 'cells'
        print('============================= ')
        print('Local MAE is %.6f' % score)
        print('=============================')

        end = time.time()

        del self.ValidData
        gc.collect()

        print('time elapsed %ds' % (end - start))