top of page

Installing TensorFlow with GPU support

The second project of Self-Driving Car program is "Traffic Sign Classifier Project". In order to gain a quicker training process, TensorFlow with GPU support need to be installed properly first.

Besides the official web instruction: https://www.tensorflow.org/install/install_windows

WARNING: after finishing

  • CUDA® Toolkit 8.0.

  • The NVIDIA drivers associated with CUDA Toolkit 8.0.

  • cuDNN v6.1.

  • GPU card with CUDA Compute Capability 3.0 or higher.

DO NOT follow "Installing with Anaconda" steps in the official instructions, read the bottom update part FIRST !

I found the following useful resources:

  • Install Tensorflow (GPU version) for Windows and Anaconda: https://www.youtube.com/watch?v=Ebo8BklTtmc&t=673s

  • [Windows] Tensorflow GPU fails to find CUDA: https://github.com/tensorflow/tensorflow/issues/5968

  • How to fix "python is not recognized as an internal or external command": https://www.youtube.com/watch?v=uXqTw5eO0Mw

You can do the test if the tensorflow take the GPU advantage using the following method:

How to tell if Jupyter notebook is using GPU: https://discussions.udacity.com/t/how-to-tell-if-jupyter-notebook-is-using-gpu/217660

==================================================================================

Issue update:

When following above steps, I did the second project Project: Build a Traffic Sign Recognition Program: https://github.com/udacity/CarND-Traffic-Sign-Classifier-Project

Everything worked fine. Everything started to collapse after I closed my Anaconda CMD, and tried to continue my project later. When I try to activate carnd-term1 environment, I got:

"

usage: conda [-h] {keygen,sign,unsign,verify,unpack,install,install-scripts,convert,version,help} ... conda: error: invalid choice: '..checkenv' (choose from 'keygen', 'sign', 'unsign', 'verify', 'unpack', 'install', 'install-scripts', 'convert', 'version', 'help')

"

I can do NOTHING. I searched online and realized this issue is pretty new, and the problem is still being discussed in "conda command failure #6171": https://github.com/ContinuumIO/anaconda-issues/issues/6171.

I finally solved my problem following this: https://stackoverflow.com/a/46493533/8936445

In short, after activating the created environment, instead of "pip install --ignore-installed --upgrade tensorflow-gpu", use "pip install tensorflow-gpu".

So, the safe steps to follow can be summarized as:

  • Finish "Requirements to run TensorFlow with GPU support" in the official install instruction: https://www.tensorflow.org/install/install_windows. Detailed steps can refer "Install Tensorflow (GPU version) for Windows and Anaconda": https://www.youtube.com/watch?v=Ebo8BklTtmc&t=673s

  • Follow all instructions of "[Windows] Tensorflow GPU fails to find CUDA": https://github.com/tensorflow/tensorflow/issues/5968

  • Activate carnd-term1, install tensorflow-gpu using "pip install tensorflow-gpu" instead of "pip install --ignore-installed --upgrade tensorflow-gpu".

A good resources about understanding environment: GPU in carnd-term1 environment (https://discussions.udacity.com/t/gpu-in-carnd-term1-environment/277322)

The "carnd-term1" environment doesn't contains spyder packag, but "IntroToTensorFlow" does. if you would like to use spyder, install tensorflow-gpu after activating "IntroToTensorFlow" environment.


Recent Posts
Archive
bottom of page