The Top-1 score obtained by testing ResNet50 with the test set is strange.

[henry81003003]

Hi authors, I have a problem while inference ResNet50 with the ImageNet test set.
When I inference ResNet50 with the ImageNet test set, I got a strange result that the Top-1 score is 0.102.
However, if I inference ResNet50 with the ImageNet train set, the Top-1 score that I got is about 75.50.
As mentioned above, it's quite strange during inference with the test set.
I've trained it for 50 epochs, but the result told me that the model didn't learn anything from training, it's guessing.
The following picture with blue values is the result I got from the train set, the other is the result I got from the test set.
Hope you can give me some advice to fix this strange situation, thanks.

The result from the train set:

The result from the test set: