Tensorflow batchsize affects precision

Question

The same input, only different batch sizes.

• Why the outputs are different?
• How to avoid the difference or force the output to be the same?

<