Difference between revisions of "Tensorflow with gpu"

Testing setup

• Supported card GeForce GTX 750 Ti (list of supported graphic cards):

~$ python3
Python 3.5.2 (default, Nov 23 2017, 16:37:01) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> hello = tf.constant('Hello, World!')  
>>> sess = tf.Session()
2018-04-26 18:14:05.427668: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning 
NUMA node zero
2018-04-26 18:14:05.428033: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1344] Found device 0 with properties: 
name: GeForce GTX 750 Ti major: 5 minor: 0 memoryClockRate(GHz): 1.1105
pciBusID: 0000:01:00.0
totalMemory: 1.95GiB freeMemory: 1.53GiB
2018-04-26 18:14:05.428061: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1423] Adding visible gpu devices: 0
2018-04-26 18:14:05.927106: I tensorflow/core/common_runtime/gpu/gpu_device.cc:911] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-04-26 18:14:05.927149: I tensorflow/core/common_runtime/gpu/gpu_device.cc:917]      0 
2018-04-26 18:14:05.927163: I tensorflow/core/common_runtime/gpu/gpu_device.cc:930] 0:   N 
2018-04-26 18:14:05.927313: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1041] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 1289 MB memory) -> physical GPU (device: 0, 
name: GeForce GTX 750 Ti, pci bus id: 0000:01:00.0, compute capability: 5.0)
>>> print(sess.run(hello))
b'Hello, World!'

• Unsupported card GeForce GT 610

~$ python3
Python 3.5.2 (default, Nov 23 2017, 16:37:01) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> hello = tf.constant('Hello, World!')                                                                                                                                                                            
>>> sess = tf.Session()                                                                                                                                                                                                 
2018-04-26 13:00:19.050625: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA                                           
2018-04-26 13:00:19.181581: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1344] Found device 0 with properties:                                                                                                             
name: GeForce GT 610 major: 2 minor: 1 memoryClockRate(GHz): 1.62                                                                                                                                                                
pciBusID: 0000:81:00.0                                                                                                                                                                                                           
totalMemory: 956.50MiB freeMemory: 631.69MiB                                                                                                                                                                                              
2018-04-26 13:00:19.181648: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1394] Ignoring visible gpu device (device: 0, name: GeForce GT 610, pci bus id: 0000:81:00.0, compute capability: 2.1) with Cuda compute capability 2.1. The minimum required Cuda capability is 3.5.                                                                                                                                                                                                              
2018-04-26 13:00:19.181669: I tensorflow/core/common_runtime/gpu/gpu_device.cc:911] Device interconnect StreamExecutor with strength 1 edge matrix:                                                                                       
2018-04-26 13:00:19.181683: I tensorflow/core/common_runtime/gpu/gpu_device.cc:917]      0                                                                                                                                                       
2018-04-26 13:00:19.181695: I tensorflow/core/common_runtime/gpu/gpu_device.cc:930] 0:   N                                                                                                                                                       
>>> print(sess.run(hello))                                                                                                                                                                                                                             
b'Hello, World!'

• As a quickfix had to install CuDNN 7.0.5 instead of latest:

https://stackoverflow.com/questions/49960132/cudnn-library-compatibility-error-after-loading-model-weights

• Print tensorflow version

>>> print(tf.__version__)

Problems

• [SOLVED] AttributeError: '_NamespacePath' object has no attribute 'sort'

# Notes:
 After updating some packages probably. python3?
# How to reproduce:
1:
~$ python3
>>> import tensorflow
2:
~$ virtualenv --system-site-packages -p python3
# Solution:
~$ sudo pip3 install setuptools --upgrade

Walkthrough for CUDA 10.1 (20190602)

Install CUDA

• In this guide there's a link to CUDA toolkit.
  • That toolkit (CUDA Toolkit 10.1 update1 (May 2019)) also updated the system driver to 418.67
  • Reboot

Install cuDNN

• Have to have an account with NVIDIA - downloaded cuDNN v7.6.1 (June 24, 2019), for CUDA 10.1

Option 1: installing tensorflow from source

Basically, this guide, some key notes:

• Install bazel - version 0.25.2 (newer will not work)
• To build, read this link:

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r1.14
./configure
 
bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
# 4-5 hours later
./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
sudo pip3 install /tmp/tensorflow_pkg/tensorflow-[Tab]

• Testing:

~$ python3
>>> import tensorflow as tf
>>> hello = tf.constant('Hello, World!')                                                                                                                                                                            
>>> sess = tf.Session()

Option 2: using docker

Follow this guide. Key notes:

• Tensorflow docker image requires nvidia docker image, nvidia docker image requires apt install nvidia-docker2, nvidia-docker2 requires apt install docker-ce:

- https://github.com/NVIDIA/nvidia-docker
- https://docs.docker.com/install/linux/docker-ce/ubuntu/

• Test run:

# Test 1: GPU support inside container:
sudo docker run --runtime=nvidia --rm nvidia/cuda:10.1-base nvidia-smi
# Test 2: Test all together
sudo docker pull tensorflow/tensorflow:latest-gpu-py3-jupyter
sudo docker run --runtime=nvidia -it --rm tensorflow/tensorflow:latest-gpu-py3-jupyter python -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))"
# Test 3: Run a local script (and include a local dir) in contatiner:
https://www.tensorflow.org/install/docker

Setup walkthrough for CUDA 10.2 (Dec 2019)

Install CUDA

• In this guide there's a link to CUDA toolkit.
  • The toolkit (CUDA Toolkit 10.2) also updated the system driver to 440.33.01
  • Will have to reboot

Docker

Instructions

https://www.tensorflow.org/install/docker

Quote:

Docker is the easiest way to enable TensorFlow GPU support on Linux since only the NVIDIA® GPU driver is required on the host machine (the NVIDIA® CUDA® Toolkit does not need to be installed).

Docker images

Where to browse: https://hub.docker.com/r/tensorflow/tensorflow/:

nvidia-docker

Somehow it was already installed.

• Check NVIDIA docker version

~$ nvidia-docker version

• In the docs it's clear that Docker version 19.03+ should use nvidia-docker2. For Docker of older versions - nvidia-docker v1 should be used.
• It's not immediately clear about the nvidia-container-runtime. nvidia-docker v1 & v2 should have already registered it.

Notes

• Can mount a local directory in a 'binding' mode - i.e., update files locally so they are updated in the docker container as well:

# this will bind-mount directory target located in $(pwd), which is a dir the command is run from 
# to /app in the docker container

~$ docker run \
   -it \
   --rm \
   --name devtest \
   -p 0.0.0.0:6006:6006 \
   --mount type=bind,source="$(pwd)"/target,target=/app \
   --gpus all \
   tensorflow/tensorflow:latest-gpu-py3 \
   bash

• How to run tensorboard from the container:

# from here
# From the running container's command line (since it was run with 'bash' in the step above).
# set a correct --logdir
root@e9efee9e3fd3:/# tensorboard --bind_all --logdir=/app/log.txt  # remove --bind_all for TF 1.15
# Then open a browser:
http://localhost:6006

Tensorflow and OpenCV building notes

Build 1

1. TF 1.15.0
2. CUDA 10.0 and Toolkit and stuff
3. OpenCV 3.4.9

TF 1.15.0

• Will build with Bazel 0.25.2 (installed from deb archive)
• TF - downloaded as tensorflow-1.15.0.tar.gz

1. Unpack
2. cd tensorflow-1.15.0
3. ./configure
4.

Build 2

1. TF 1.13.1
2. CUDA 10.0 and Toolkit and stuff
3. OpenCV 3.4.9

TF 1.13.1