Deep Learning in QGIS: installing Tensorflow

Question

I am trying to install tensorflow (with or without GPU support) with the keras API in the QGIS 3.4 LTR python 3 environment but without success. I do have a working install of tensorflow-gpu that I use in anaconda. I also have a working install of scikit-learn (and others) in the QGIS 3 env. So I think my method using the OSgeo4 shell and pip seems mostly OK. So from the python console in QGIS I can type:

import sklearn

And that works fine. But the same thing is not true for tensorflow and keras. First I tried to install tensorflow-gpu, when using pip I got no error messages. But:

import tensorflow as tf

returns this error:

ImportError: Could not find 'cudart64_100.dll'. TensorFlow requires that this DLL be installed in a directory that is named in your %PATH% environment variable. Download and install CUDA 10.0 from this URL: https://developer.nvidia.com/cuda-90-download-archive

I have used Settings>options>System and added all the CUDA paths to my environment variables. These do point to the folder where cudart64_100.dll is found.

If I un-install tensorflow-gpu and downgrade to the cpu version, I get another problem after an attempted import:

AttributeError: type object 'h5py.h5.H5PYConfig' has no attribute '__reduce_cython__'

This seems to be related to v 2.10 of h5py.

Ultimately, is there a solution here? Is it even possible to install tensorflow in QGIS and use it directly from the PyQGIS console?

• Patrice Carbonneau: I have made progress and am now running my code with tensorflow, non-GPU. The solution was to downgrade h5py to v 2.9. But the environment path bug for tensorflow-gpu and CUDA tools persists.

Answer 1

Solved my problem. For reference:

1- pip install tensorflow-gpu in the Python 3.7 env of QGIS, be sure to use the py3_env command.

2-Be sure h5py is v2.9. I had to force this. This will prevent tensorflow bugs and QGIS crashes

3- In QGIS, use Settings>Options>System. Tick 'use custom variables'

4- Pulldown the small menu to Append

5- enter PATH as the variable

6- Find the location of your CUDA folder and the ...\bin and the ...\libnvvp paths

7- add these paths with a ; separator as the value of the PATH variable.

This will give you a working install of TF GPU that can be called from the QGIS python console!