1. Install CUDA 11.0
https://developer.nvidia.com/cuda-downloads
2. Install CUDNN 8.0.3 for CUDA 11.0
(Make sure to install the correct NVIDIA drivers as well)
https://docs.nvidia.com/deeplearning/cudnn/install-guide/index.html
3. Install keras version 2.4.1
pip install keras==2.4.1
4. Uninstall tensorflow
pip uninstall tensorflow
5. Install tf-nightly version 924
pip install tf-nightly==2.4.0.dev20200924
(If you get error due to ‘wrapt’ follow this:
conda update --force conda
conda update --all
conda update wrapt)

6. Install tf-nightly-gpu version 924
pip install tf-nightly-gpu==2.4.0.dev20200924

7. Install tf-estimator-nightly version 924
pip uninstall tf-estimator-nightly
pip uninstall tensorflow-estimator
pip install tf-estimator-nightly==2.4.0.dev2020092401

8. Make sure that that the CUDA_PATH environment variable is set to …\CUDA\v11.0
9. Restart the system
After all these steps, The following code should produce this:


import tensorflow as tf
from tensorflow.python.platform import build_info as tf_build_info
print(tf_build_info.build_info)
print(tf.__version__)
print("Num GPUs Available: ", len(tf.config.experimental.list_physical_devices('GPU')))
------
OrderedDict([('cpu_compiler', 'C:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.27.29110/bin/HostX64/x64/cl.exe'), ('cuda_compute_capabilities', ['sm_35', 'sm_50', 'sm_60', 'sm_70', 'sm_75', 'compute_80']), ('cuda_version', '64_110'), ('cudart_dll_name', 'cudart64_110.dll'), ('cudnn_dll_name', 'cudnn64_8.dll'), ('cudnn_version', '64_8'), ('is_cuda_build', True), ('is_rocm_build', False), ('msvcp_dll_names', 'msvcp140.dll,msvcp140_1.dll'), ('nvcuda_dll_name', 'nvcuda.dll')])
2.4.0-dev20200924
Num GPUs Available:  1



Other packages:

1. Install tensorflow privacy (https://github.com/tensorflow/privacy)
pip install tensorflow_privacy