如何使用Python在Windows上正确安装Tensorflow？

Question

I'm trying to use tensorflow with my PC's GPU (Nvidia RTX 3070Ti) in python-conda environment. I'm solving a small image-classification problem from kaggle. I've solved it in google-collab, but now I'm intrested in solving it on my local machine. However TF doesn't work properly locally and I have no idea why. I've read tons of solutions but it didn't help yet.

I'm following this guide and always install proper versions of TF and CUDA: https://www.tensorflow.org/install/source_windows

cuda-toolkit 10.1, cudnn 7.6, tf-gpu 2.3, python 3.8

Also I've installed latest NVidia drivers for videocard.

What I've tried:

1. I've installed proper version CUDA-toolkit and CUDnn from nvidia site. I've installed it properly and included everything that was needed into PATH. I've checked it - MS Visiual Studio finds both CUDA and CUDnn and can work with it. I've installed proper version of Tensorflow-GPU using conda into my environment.

Result: TF can't find my GPU and uses only CPU.

2. I've removed all CUDA and CUDAnn drivers. I've installed CUDA-toolkit, CUDnn and Tensorflow-GPU python packages into my conda environment.

Result: TF recognizes my GPU and uses it: But during DNN training happens error: Failed to launch ptxas Relying on driver to perform ptx compilation. Modify $PATH to customize ptxas location. Failed to launch ptxas Relying on driver to perform ptx compilation. Modify $PATH to customize ptxas location. And training goes very bad - accuracy is very low and doesn't improving.

When I use absolutely same code and data on google-collab, everything is going smoothly - I get ~90% accuracy on 5th epoch.

3. I've tried tf 2.1 and relevant cuda and cudnn, but it's still same result!

4. I've tried to install cudatoolkit-dev, but it didn't help to solve ptxas problem.

I'm about to give up and use PyTorch instead of Tensorflow.

Answer 1

So here is what worked for me:

1. Create 3.9 python environment
2. Install cuda and tensorflow packages from "Esri":

 conda install -c esri cudatoolkit conda install -c esri cudnn conda install -c esri tensorflow-gpu

3. Then install tensorflow-hub:

 conda install -c conda-forge tensorflow-hub

It will downgrade installations from previous steps, but it works. Maybe installing tensorflow-hub first could help to avoid it, but I didn't test it.