Removing the re-size for validation data, which breaking the validation accuracy of CIFAR training

[juliusshufan]

Description

The dataloader for example/image-classification resizes the imagesize of validation dataset to 256 by default, and which breaks the validation accuracy of CIFAR10, this PR trying to fix it.
Below firgure is the validation accuracy trends without this fix (model is ResNet-50).

Checklist

Changes

example/image-classification/common/data.py

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

[juliusshufan]

@szha may I have your comments, I use this example for the regression tracking of mxnet with mkldnn backend, I'll be helpful if this issue can be addressed.
Thanks.

[zhreshold]

Does it affect the imagenet validation perf?

[juliusshufan]

@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.

[TaoLv]

Seems this regression also goes into 1.3.0rc0 realese.

[juliusshufan]

@zhreshold sorry for pushing, may I have you comments or approval?

Thanks.

[zhreshold]

@juliusshufan Thanks for the result, it's fair enough to go through. Again, thanks for the contribution!