Setting_up_TensorFlow_Conda_Environment/Install TensorFlow under Conda Env on Win 10 or 11.md

In John-R-Wallace-NOAA/FishNIRS: NIRS on fish structures.

 
# -- Install TensorFlow and CUDA in Conda under Win 10 & 11; GPU works in Python and in R --
#        This install uses a mix of conda-forge and pip package managers.

# To start, you first need to see your system's NVIDIA graphics card as a GPU under Task Manager on Windows 10 or 11:
# That should not be a problem in Windows 10 or 11, see the link below for some general info, but use the install code below.
#      https://hackmd.io/@husohome/Byb6kP6WP

# For a picture see here:
#     https://towardsdatascience.com/setting-up-tensorflow-gpu-with-cuda-and-anaconda-onwindows-2ee9c39b5c44

# but don't follow that site past installing the Community Verson of MS Visual Studio, if needed:
#     https://visualstudio.microsoft.com/vs/community/

# The Anaconda software is here:
#    https://www.anaconda.com/products/distribution

# After the Anaconda installation follow the steps below.

# In the Anaconda Powershell Prompt window
conda create -n tf python==3.8 conda=4.8

conda env list

conda activate tf

conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0

# Check CUDA installation 
nvcc --version
conda list cuda


# Anything above 2.10.* is not supported on the GPU on Windows Native
pip install "tensorflow<2.11"

# Find version number of tensorflow (2.10.1)
conda list tensorflow

# Verfiy TensorFlow - single-line approach
python -c "import tensorflow as tf; print('\n===== \n GPU Devices: ',tf.config.list_physical_devices('GPU'), '\n=====\n')"
python -c "import tensorflow as tf; print('\n\n=====\n', tf.reduce_sum(tf.random.normal([1000, 1000])), '\n======\n' )"

# Verfiy TensorFlow - interactive approach
python
>>> 
import tensorflow as tf
print(tf.__version__)
print(tf)

tf.config.list_physical_devices('CPU')
tf.config.list_physical_devices('GPU')
len(tf.config.list_physical_devices('GPU'))

a = tf.constant(7)
b = tf.constant(10)
print(tf.add(a,b))

print(tf.reduce_sum(tf.random.normal([1000, 1000])))

quit()
>>>


# ------ Extra testing using Jupyter Notebook -----

# Bex T. in the the link below does not install the CUDA and cuDNN under Conda, but below is a quote and 
#   a Jupyter Notebook test:
#      https://towardsdatascience.com/how-to-finally-install-tensorflow-gpu-on-windows-10-63527910f255

# "Think of cuDNN (NVIDIA CUDA® Deep Neural Network library) as a library for Deep Learning 
#     using CUDA and CUDA as a way to talk to the GPU."

pip install jupyterlab ipykernel

ipython kernel install --user --name=tf

# Use ctrl-c to end
mkdir jupyter_folder
jupyter-notebook --notebook-dir jupyter_folder

# Create a new notebook and Run (from the tab above) this 4-line snippet:
import tensorflow as tf
from tensorflow.python.client import device_lib

print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))
device_lib.list_local_devices()


# Back in the Anaconda Powershell Prompt
ctrl-c

conda deactivate


# ----- The installs above do not include the optional TensorRT (not for lack of trying under conda on native Windows). -----

# What is TensorRT?
#      https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tensorrt

John-R-Wallace-NOAA/FishNIRS documentation built on April 12, 2025, 12:59 a.m.