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:
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.
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. 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.
I've tried tf 2.1 and relevant cuda and cudnn, but it's still same result!
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.
So here is what worked for me:
Create 3.9 python environment
Install cuda and tensorflow packages from "Esri":
conda install -c esri cudatoolkit
conda install -c esri cudnn
conda install -c esri tensorflow-gpu
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.
Related
This weekend I have been trying a lot to install and get Tensorflow with GPU support to work on my computer, but I am not very experienced in using pip/conda and are now quite confused after watching and trying a lot of different tutorials/approaches from the web.
I have a GeForce GTX 1650 graphics card, and I have installed Cuda 10.0 (also 11.2, but I removed it from "PATH" and are only using the 10.0 version, I don't think that's a problem).
I have downloaded cuDNN 7.5.0 for CUDA 10, and I think that I have copied and placed the files correctly (installed cuDNN).
I am just trying to get some version of Tensorflow-gpu to work, but you can see the Tensorflow version i have been trying for now on the image.
I have tried to install and uninstall Python from my computer (I've also reinstalled Anaconda a lot of times), because I am not sure if I need to have a Python version installed (on my system) if I install a version of Python inside my Anaconda environment (in my example Python 3.7).
Does anyone know how to install Tensorflow GPU on Windows 10 with my settings (cuDNN 7.5.0, CUDA 10), or maybe have encountered some trouble with Python versions or Anaconda problems similar to mine?
Follow these steps to install Tensorflow GPU on windows system.
Make sure right version of Visual studio is installed. Check here.
Follow the instructions mentioned here to setup CUDA for windows system
Install Tensorflow
#check current python version
python --version
#Create the virtual environment
conda create -n tf python=PYTHON_VERSION
#Activate the tf environment
conda activate tf
#Install Tensorflow
pip install tensorflow
#Install CUDA and cuDNN using conda and make sure CUDA and cuDNN version should match the Tensorflow version
conda install -c anaconda cudatoolkit=10.0 cudnn=7.5
I want to use Tensorflow on GPU. So I install all the needed tool and installed as below-
CUDA-11.2
CUDNN-11.1
Anaconda-2020.11
Tensorflow-GPU-2.3.0
I tested that my cuda,cudnn is working using deviseQuery example.
But Tensorflow not used GPU. Then i find that version compatibility issue is possible so i innstalled CudaToolkit,cudnn using conda environment checking with version compatibility on Tensorflow website which is given below.
CUDA-10.2.89
CUDNN-7.6.5
Tensorflow-GPU-2.3.0
But after this try Tensorflow-GPU not used GPU,yet. so what i am doing now? Any steps or suggestion require.
The installation engine has a problem for tensorflow-gpu 2.3 in Anaconda on Windows 10.
Workaround is to explicitly specify the correct tensorflow build:
conda install tensorflow-gpu=2.3 tensorflow=2.3=mkl_py38h1fcfbd6_0
I want to run the project using Anaconda, TensorFlow 2.3, Keras 2.4.3 (CNN example). OS Windows 10.
I installed Visual Studio 2019 Community Edition, CUDA 10.1 and cudnn 8.0.5 for CUDA 10.1.
Using Anaconda I created an environment with TensorFlow (tensorflow-gpu didn't help), Keras, matplotlib, scikit-learn. I tried to run it on CPU but it takes a lot of time (20 minutes for just 1 epoch when there are 35).
I need to run it using GPU, but TensorFlow doesn't see my GPU device (GeForce GTX 1060). Can someone help me find the problem? I tried to solve the problem using this guide tensorflow but it didn't help me.
This works 100%, no need to install anything manually (cuda for example)
conda create --name tf_gpu tensorflow-gpu
Ok so I tried to install all the components into new anaconda environment. But instead of "conda install tensorflow-gpu" I decided to write "pip install tensorflow-gpu" and now it works via GPU...
Just a heads up, the Cudnn version you were trying to use was incompatible.
Listing Versions and compatible CUDA+Cudnn
You can go here and then scroll down to the bottom to see what versions of CUDA and Cudnn were used to build TensorFlow.
Anaconda has different packages for Tensorflow with and without GPU support.
In particular, to install Tensorflow with GPU, you should run:
conda install tensorflow-gpu
While for the non-GPU version, you should install:
conda install tensorflow
By checking the version of the installed package, conda installs Tensorflow version 2.1.
But as of today the latest version of Tensorflow is 2.3. Furthermore, as can be seen in the Tensorflow officla documentation, the latest version can be installed with
pip install tensorflow
This package is said in the documentation to be good both for CPU and GPU versions of Tensorflow. Moreover, the documentation states that the packages for CPU and GPU were different for "for releases 1.15 and older".
Why Anaconda provides 2.1 in two different packages, given that the package should be the same for any version > 1.15?
Which one should I install, the pip version or the conda version? An article in Anaconda blog specifies that the version provided with conda is faster, but the article is old (2018) and refers to an old version of Tensorflow (1.10)
By checking the version of the installed package, conda installs Tensorflow version 2.1.
But as of today the latest version of Tensorflow is 2.3. Furthermore
That is only because you are (probably?) on windows. As you can see here tensorflow is available as 2.3 from conda default channels, but currently only on linux.
The reason is also stated on the website you have linked (emphasis mine):
Anaconda is proud of our efforts to deliver a simpler, faster experience using the excellent TensorFlow library. It takes significant time and effort to add support for the many platforms used in production, and to ensure that the accelerated code is still stable and mathematically correct. As a result, our TensorFlow packages may not be available concurrently with the official TensorFlow wheels. We are, however, committed to maintaining our TensorFlow packages, and work to have updates available as soon as we can.
In short: The Anaconda team is creating custom builds of tf against the intel mkl library to speed up calculations on the CPU. Earlier on the same website they also mention that they create builds for different cuda versions.
Why Anaconda provides 2.1 in two different packages, given that the package should be the same for any version > 1.15?
The tensorflow-gpu package is only a meta-package, i.e. it is only used to install a different build of tensorflow with different dependencies (also enabling you to install for different cuda versions). The official releases only allow for combinations of tensorflow version and cuda.
Which one should I install, the pip version or the conda version? An article in Anaconda blog specifies that the version provided with conda is faster, but the article is old (2018) and refers to an old version of Tensorflow (1.10)
Reading said article, the speed up is linked to building against the intel mkl library, which speeds up calculations on the CPU. Given that for your setup, you can only get tensorflow 2.1 installed when using conda, you will need to ask yourself if you rely on the newest tensorflow version and if you don't need the accelerated cpu code. There is usually nothing wrong with installing the newest tensorflow using pip. Just make sure that you create a new environment for said tensorflow version and only install/update tensorflow or any of its dependencies using pip in that environment. There is general advice to not mix conda and pip installations too much, since one could break the other (since they are using different ways to resolve dependencies), but you should be fine when using a seperate env
If you are using Anaconda then you can use conda to install tensorflow. For the cpu version enter
conda install tensorflow
for the gpu version enter
conda install tensorflow-gpu.
If you are using Windows it will install version 2.1.0, the cuda toolkit version 10.1.243 and cudnn version 7.6.5. Note conda can only install tensorflow up to version 2.1.0 on Windows operating system. If you want tensorflow 2.2.0 or 2.3.0 install it with pip using pip after you have installed 2.1. The cuda toolkit and cudnn work with version 2.2 and 2.3. One other thing. Use python3.7 not 3.8. Apparently when you install tensorflow with conda it will not work with 3.8.
If you use pip to install tensorflow 2.1 or higher it includes both the cpu and gpu versions however you have to go through a manual processes to install the Cuda Toolkit and cudnn. This includes downloading the files from NVIDIA . You also have to change your PATH environmental variable.
I have some problem with tensorFlow. I'm trying to install it with GPU on my manjaro linux with GTX 1060.
When I try to import tensorFlow in python with:
import tensorflow as tf
I get this error:
{...} ImportError: libcublas.so.8.0: cannot open shared object file:
No such file or directory {...}
With pip, I have installed tensorFlow-gpu:sudo pip install tensorflow-gpu
When I try to install cuda-8.0 (with pacaur -Syu cuda-8.0), after a very long loading, I got an error. Now when I try to install it, it does this:
Errors occurred, no packages were upgraded
Even if it's not on my pacaur list, and there is no reinstalling signed
I have install Keras with: sudo pip install Keras
I have install cudNN with: pacaur -Syu cudnn
I have installed my nvidia driver with (if I remember it right):pacaur -Syu nvidia
I am not familiar with manjaro. Assume you wanna install TensorFlow 1.4, the order would be:
Install latest Nvidia driver (version 384.xx or higher). Check its status in a terminal with nvidia-smi.
Install CUDA 8.0 without the GPU driver (as you have done it in step 1).
Add PATH=/usr/local/cuda-8.0/bin to the environment (in Ubuntu it's /etc/environment).
Added driver and CUDA paths to LD_LIBRARY_PATH. In Ubuntu, it is done by adding export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:/usr/local/cuda/lib64:/usr/lib/nvidia-384:/usr/local/cuda/extras/CUPTI/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} to /etc/bash.bashrc. At this point, you should be able to check CUDA version by nvcc --version.
Copy CUDNN files to somewhere and add that path to LD_LIBRARY_PATH. CUDNN needs no installation.
Install TensorFlow 1.4.
If you wanna install other versions of TensorFlow, you need to first check the supported versions of CUDA and CUDNN.
Hope this helps.