Tensorflow 1.7 install


(Peter) #1

Hello,
I am following with great interest the various lessons.

Now, I would like to do some stuff on my own and need also Tensorflow (possibly the latest 1.7 version with GPU support installed in the fastai anaconda environment).

I tried to simply do the following command line
pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0-cp36-cp36m-linux_x86_64.whl
Possible has to do with some drivers on which I don’t want to start messing up the default fastai installation

but it gave me problems.
It would be good if ths library would be part of the standard fastai install script.

Thanks for considering
Peter


(Ramesh Sampath) #2

FastAI Library doesn’t use TensorFlow, so there’s no reason to include it. If you are having trouble installing it, you might get better help if you post this in StackOverflow.

Also, I would say create two different conda environments. One for fastai and one for your Tensorflow experiments. That way it might be easier to keep them in sync. But if your usecase is to use some functions from both, you may need to combine and have one script. Most folks here in FastAI community don’t deal with this problem.


(Lalit A Patel) #3

@peter https://www.udemy.com/complete-guide-to-tensorflow-for-deep-learning-with-python/ seems to be a good TensorFlow course (from a popular teacher).

However, I am not sure whether that course handles computer vision and other stuff that will need GPU.


(Agis) #4

Hi @peter, I don’t know if it’s going to help you but an alternative would be to use docker images with everything pre-setup.

You just need to install docker and then pull the image you want and run it

docker pull tensorflow/tensorflow should give you the latest offical tensorflow image and you can find instructions [here] (https://hub.docker.com/r/tensorflow/tensorflow/).

Alternatively, if you want my own images which have a big bunch of libraries preinstalled try:

docker pull agispof/research:tensorflow

you can get gpu ready images as well but you need to have nvidia-docker installed as well to get them running and you would need to get the gpu variants, i.e. run

docker pull tensorflow/tensorflow:latest-gpu for the official the tensorflow one, and/or
docker pull agispof/research:tensorflow-gpu for mine.

hopes this helps


#5

I did conda install tensorflow-gpu and got the latest tensorflow installed and working with no problems. You may need to also update to the latest CUDA (9.0, 9.1 or 9.2) and remove the old versions of CUDA though.