TF.Keras Load Model predict low accuracy when using ROCm

[jszhtian]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 20.04.1 LTS
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: No
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): tensorflow-rocm 2.3.1
• Python version: 3.8
• Bazel version (if compiling from source): N/A
• GCC/Compiler version (if compiling from source): N/A
• CUDA/cuDNN version: ROCm
• GPU model and memory: RX480 8GB

Describe the current behavior
I want to test the model accuracy after traning the model on a Computer Cluster.
When I use ROCm version TensorFlow on Linux, It will give me a extreme low accuracy compare with use pure CPU version Tensorflow on Windows.

Describe the expected behavior
Almost the same accuracy compare with pure CPU code.
Standalone code to reproduce the issue

TF2ROCm.zip

[xuhuisheng]

Maybe same problem
#1106

[ielenik]

Duplicate of #1105