Why is there so much difference between the accuracy of model prediction and training accuracy?

Question

Here is my code:

dataset: mnist;

model: resnet

based on tensorflow 1.15.0

During resnet model training, I noticed this behavior in my experiment. The