xiaochus / TrafficFlowPrediction

Traffic Flow Prediction with Neural Networks(SAEs、LSTM、GRU).

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

训练过程中mape值超高

zydarChen opened this issue · comments

你好,想问一下训练的过程中mape值超高是什么原因?在每epoch训练的时候,mape会突然飙升,虽然对最终结果没有影响,但是很好奇。

Epoch 110/600
 256/7375 [>.............................] - ETA: 1s - loss: 0.0025 - mean_absolute_percentage_error: 21.2901
 768/7375 [==>...........................] - ETA: 1s - loss: 0.0026 - mean_absolute_percentage_error: 22.3736
1280/7375 [====>.........................] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 22.8402
1792/7375 [======>.......................] - ETA: 0s - loss: 0.0025 - mean_absolute_percentage_error: 8754.2923
2304/7375 [========>.....................] - ETA: 0s - loss: 0.0025 - mean_absolute_percentage_error: 6813.5487
2816/7375 [==========>...................] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 8774.6254
3328/7375 [============>.................] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 14578.9923
3840/7375 [==============>...............] - ETA: 0s - loss: 0.0025 - mean_absolute_percentage_error: 12638.1525
4352/7375 [================>.............] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 33742.5767
4864/7375 [==================>...........] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 30193.0208
5376/7375 [====================>.........] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 27320.6417
5888/7375 [======================>.......] - ETA: 0s - loss: 0.0026 - mean_absolute_percentage_error: 24946.6725
6400/7375 [=========================>....] - ETA: 0s - loss: 0.0025 - mean_absolute_percentage_error: 22952.6749
6912/7375 [===========================>..] - ETA: 0s - loss: 0.0025 - mean_absolute_percentage_error: 21253.7816
7375/7375 [==============================] - 1s 170us/step - loss: 0.0026 - mean_absolute_percentage_error: 19921.1577 - val_loss: 0.0033 - val_mean_absolute_percentage_error: 17.6413
commented

@zydarChen 根据mape的公式,计算的时候会除以ground truth,所以当ground truth为0时,为了防止除0错误会除以一个极小值导致mape很高。

OK,明白啦,谢谢