can not reproduce the result of book/03.image_classification/

[NHZlX]

Model: resnet32
dataset: cifar10
using https://github.com/PaddlePaddle/book/blob/develop/03.image_classification/train.py
This looks like overfitting, that the test results are much worse than the training accuracy...
here is the log:
..........................................................................................
Test with Pass 12, score, 1.020731, {'classification_error_evaluator': 0.24729999899864197}

Pass 13, Batch 0, Cost 0.191999, {'classification_error_evaluator': 0.0390625}
5.97 seconds 100 batch
...................................................................................................
Pass 13, Batch 100, Cost 0.157929, {'classification_error_evaluator': 0.0703125}
78.67 seconds 100 batch
...................................................................................................
Pass 13, Batch 200, Cost 0.251160, {'classification_error_evaluator': 0.0859375}
77.86 seconds 100 batch
...................................................................................................
Pass 13, Batch 300, Cost 0.143929, {'classification_error_evaluator': 0.0546875}
58.15 seconds 100 batch

[qingqing01]

• The momentum and decay_rate does not work according to issue The hyper parameters of paddle.optimizer does not work in v2 API. #2042 .

• Need to check data argumentation, like flipping. Some preprocess can refer to the v1 version.

@NHZlX Thanks for your work.

[JiayiFeng]

Close this issue due to inactivity. please feel free to reopen it if more information is available.