[GitHub] [incubator-mxnet] juliusshufan commented on issue #12362: Removing the re-size for validation data, which breaking the validation accuracy of CIFAR training

Sun, 09 Sep 2018 20:27:02 -0700 

@zhreshold Thanks for the feedback, I use the pre-trained ResNet-50 model and 
validation-set of ImageNet-1k, and the accuracy with/w.o my changes are same: 
see test result:

**Without** my changes:
INFO:root:Finished with 126.286497 images per second
INFO:root:('accuracy', 0.753156969309463)
INFO:root:('top_k_accuracy_5', 0.9257512787723785)

**With** my changes:
INFO:root:Finished with 126.029153 images per second
INFO:root:('accuracy', 0.753156969309463)
INFO:root:('top_k_accuracy_5', 0.9257512787723785)

The imagenet validation perf and accuracy are **same**.

But without my changes, the validation accuracy trends of 
CIFAR10+ResNet50/VGG16 are as below, obviously it is not expected.

![image](https://user-images.githubusercontent.com/33112206/45274887-02611b00-b4ec-11e8-84ff-1dd02c93cff4.png)
![image](https://user-images.githubusercontent.com/33112206/45274907-2a507e80-b4ec-11e8-97c8-e3208d8dc0b0.png)


[ Full content available at: 
https://github.com/apache/incubator-mxnet/pull/12362 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to