High Training Accuracy But Low Prediction Performance For Tensorflow's Official Mnist Model

I'm new to machine learning and I was following along with the Tensorflow official MNIST model (https://github.com/tensorflow/models/tree/master/official/mnist). After training the

Solution 1:

The MNIST images are white-on-black; the images you've linked are black-on-white.

Unless there's a conversion step I missed, you'll want to invert the colors before attempting detection.

Solution 2:

The MNIST images pixel value range from 0.00 to 1.00 from black to white. Usually when you use your own image it would be 0 to 255. I'm not sure if I've missed any coding that map it back down to 0.0-1.0 .Can you try printing out the array that is storing your pixel value and make sure white pixels are close to or equal to 1.0 and black pixels are close to or equal to 0.0.