Deep Learning Frameworks

See also: HPC and HTC tutorial | Run MPI On Grid'5000 | GPUs on Grid5000 | GNU Parallel | Modules | Conda | Guix | Singularity | Deep Learning Frameworks

	Note
	This page is actively maintained by the Grid'5000 team. If you encounter problems, please report them (see the Support page). Additionally, as it is a wiki page, you are free to make minor corrections yourself if needed. If you would like to suggest a more fundamental change, please contact the Grid'5000 team.

This page describes installation steps of common Deep Learning frameworks.

Deep learning with Nvidia GPUs on x86_64 nodes (common case)

conda will be used to install the frameworks (pip could be used much the same way). Installation is performed under your home directory.

Please refer to Conda's documentation on Grid'5000.

Reserve some GPU nodes with OAR

• Reserve a node with some GPUs (see the Hardware page for the list of sites and clusters with GPUs).

For instance, to reserve one GPU using OAR:

frontal:

oarsub -I -l gpu=1

Remember to add -q production option if you want to reserve a GPU from Nancy or Rennes "production" resources.

Please try to not reserve a single GPU on nodes with many GPUs (e.g. ≥ 4) if you only need to execute code on one GPU. For instance, using the gemini cluster is not very welcome for a user to use only one GPU at a time.

To reserve the full node (with all its GPUs):

frontal:

oarsub -I -l host=1

To reserve a gpu or a full node on a specific cluster, add to the oarsub command: -p cluster=<clustername>

• Once connected to the node, check GPU presence and the available CUDA version:

node:

nvidia-smi

 
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03    Driver Version: 460.91.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
(...)

Which machine should be usedd to create Conda environment?

Installing Conda packages can be time and resources consuming. You should preferably use a node (instead of a frontend) to perform such an operation. Indeed, frontend might not have enough RAM for conda.

NVIDIA Tools

NVIDIA libraries are available via Conda. It gives you the possibility to manage project specific versions of the NVIDIA CUDA Toolkit, NCCL, and cuDNN. NVIDIA actually maintains their own Conda channel. The versions of CUDA Toolkit available from the default channels are the same as those you will find on the NVIDIA channel.

node:

module load conda conda create --name NvidiaTools conda activate NvidiaTools

• To compare build numbers version from default and nvidia channel

node:

conda search --channel nvidia cudatoolkit

See:

• Nvidia doc: Installing CUDA Using Conda
• “Best practices” Managing CUDA dependencies with Conda

Cudatoolkit

• Install cudatoolkit from nvidia channel.

node:

conda install cudatoolkit -c nvidia

Note: do not forget to create a dedicated environment before.

Cuda

cuda is available in both conda-forge or nvidia channels.

• Install cuda from nvidia channel:

node:

conda install cuda -c nvidia

Note: do not forget to create a dedicated environment before.

• Installing Previous CUDA Releases

All Conda packages released under a specific CUDA version are labeled with that release version. To install a previous version, include that label in the install command to ensure that all cuda dependencies come from the wanted CUDA version. For instance, if you want to install cuda 11.3.0:

node:

conda install cuda -c nvidia/label/cuda-11.3.0

• To display the version of Nvidia cuda compiler installed:

node:

nvcc --version

PyTorch

PyTorch is an optimized tensor library for deep learning using GPUs and CPUs. It can automatically detect GPU availability at run-time.

Installation

• Load conda and activate your PyTorch environment

node:

module load conda conda create --name PyTorch conda activate PyTorch

• Simple PyTorch installation from nvidia channel

node:

conda install pytorch -c nvidia

• Custom PyTorch installation : Go on PyTorch website to see the installation command that suits you.

For instance (as of April 2023), for a full installation, you might want to combine for Linux, Pytorch Stable with Python language and specific Cuda version (e.g., 11.7). This can be done by this command:

node:

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

	Warning
	You must adapt the version number of pytorch-cuda according to your version of cuda installed on your system. GPU will not be detected by PyTorch if the version of cuda mismatches with the one installed on your system.

Verify your installation

• Check which Python binary is used:

node:

which python

/home/login/.conda/envs/env_name/bin/python

• Construct a randomly initialized tensor.

node:

python

>>> import torch
>>> x = torch.rand(5, 3)
>>> print(x)
tensor([[0.3485, 0.6268, 0.8004],
        [0.3265, 0.9763, 0.5085],
        [0.6087, 0.6940, 0.8929],
        [0.2143, 0.6307, 0.5182],
        [0.0076, 0.6455, 0.5223]])

• Print the Cuda version

node:

python

>>> import torch
>>> print("Pytorch CUDA Version is ", torch.version.cuda)
Pytorch CUDA Version is 11.6

Verify your installation on a GPU node

• Reserve only one GPU (with the associated CPU cores and share of memory) in interactive mode:

frontal:

oarsub -l gpu=1 -I

• Load conda and activate your Pytorch environment on the node

gpunode:

module load conda conda activate PyTorch

• Launch python and execute the following code:

gpunode:

python

>>> import torch
>>> print("Whether CUDA is supported by our system: ", torch.cuda.is_available())
Whether CUDA is supported by our system:  True

• To know the CUDA device ID and name of the device, you can run:

gpunode:

python

>>> import torch
>>> Cuda_id = torch.cuda.current_device()
>>> print("CUDA Device ID: ", torch.cuda.current_device())
CUDA Device ID:  0
>>> print("Name of the current CUDA Device: ", torch.cuda.get_device_name(Cuda_id))
Name of the current CUDA Device:  GeForce GTX 1080 Ti

Tensorflow

TensorFlow offers multiple levels of abstraction so you can choose the right one for your needs. Build and train models by using the high-level Keras API, which makes getting started with TensorFlow and machine learning easy.

Installation

	Warning
	By default conda install the current release of CPU-only TensorFlow, to install GPU TensorFlow use tensorflow-gpu package name. For using TensorFlow with a GPU, refer to the TensorFlow documentation on the topic, specifically the section on device placement.

• Load conda and activate a specific TensorFlow environment (see Conda Documentation)

node:

module load conda conda create --name TensorFlow conda activate TensorFlow

• Install TensorFlow from conda-forge channel

node:

conda install tensorflow-gpu

Verify your installation

• Launch python

node:

python

Python 3.10.9 (main, Jan 11 2023, 15:21:40) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>>

Verify your installation on a GPU node

• Reserve only one GPU (with the associated CPU cores and share of memory) in interactive mode:

frontal:

oarsub -l gpu=1 -I

• Load conda and activate your TensorFlow environment

gpunode:

module load conda conda activate TensorFlow

• Test the installation

gpunode:

python

>>> import tensorflow as tf
>>> print('tensorflow version', tf.__version__)
tensorflow version 2.0.0-rc1
Whether CUDA is supported by our system:  True
>>> x = [[2.]]
>>> print('hello, {}'.format(tf.matmul(x, x)))
hello, [[4.]]

To go further : https://docs.anaconda.com/anaconda/user-guide/tasks/tensorflow/

If you need TensorFlow v1, see https://www.tensorflow.org/guide/migrate

	Note
	As alternative to conda installation go on Tensorflow website to see the installation commands with pip or docker.

Keras

Keras is a high-level neural networks API, written in python, which is used as a wrapper of TensorFlow. It was developed with a focus on enabling fast experimentation. It's the recommended tool for beginners and even advanced users who don't want to deal and spend too much time with the complexity of low-level libraries as TensorFlow.

Installation

• Since version 2.4, Keras refocus exclusively on the TensorFlow implementation of Keras. Therefore, to use Keras, you will need to have the TensorFlow package installed:

node:

conda install -c conda-forge tensorflow-gpu

Note: do not forget to create a dedicated environment before.

Verify the installation

• Check which Python binary is used:

node:

which python

/home/login/.conda/envs/env_name/bin/python

• Print the Keras version

node:

python

>>> from tensorflow import keras
>>> print(keras.__version__)
2.10.0

To go further:

• Keras exemples
• Keras blog

Scikit-learn

Scikit-learn is an open source machine learning library that supports supervised and unsupervised learning. It also provides various tools for model fitting, data preprocessing, model selection, model evaluation, and many other utilities.

Installation

node:

conda install -c conda-forge scikit-learn

Note: do not forget to create a dedicated environment before.

Verify your installation

node:

python

>>> import sklearn
>>> sklearn.show_versions()
System:
    python: 3.10.9 (main, Mar  1 2023, 18:23:06) [GCC 11.2.0]
executable: /home/xxxx/.conda/envs/test/bin/python
   machine: Linux-5.10.0-21-amd64-x86_64-with-glibc2.31

Python dependencies:
          pip: 22.3.1
   setuptools: 65.6.3
      sklearn: 1.0.2
        numpy: 1.23.5
        scipy: 1.8.1
       Cython: None
       pandas: None
   matplotlib: None
       joblib: 1.2.0
threadpoolctl: 3.1.0

Built with OpenMP: True

To go further:

• scikit-learn Installation
• scikit-learn.org Tutorials
• Dataquest Scikit-learn Tutorial
• Another Python SciKit Learn Tutorial

Additional resources

• If you want to load a module in a non-interactive job, see Modules#Using_modules_in_jobs
• An in-depth tutorial contributed by a Grid'5000 user, Ismael Bada
• Many Docker images exist with ready-to-use Deep Learning software stack. They can be executed using Docker or Singularity tools (using appropriate options to enable GPU usage). See wiki pages to learn how to use these tools in Grid'5000.
• If you want to use virtualenv to manage your Python packages, it is available in Grid'5000 standard environments. Create your environment with python3 -m venv path/to/env_directory and activate it using source path/to/env_directory/bin/activate before using pip and installing packages.
• If you prefer to use conda to manage your Python packages, it is available in Grid'5000 as a module. Just execute module load conda" from a node or a frontend to make it available (Consult specific documentation of conda on Grid'5000)

Deep learning with AMD GPUs

conda will be used to install the frameworks (pip could be used much the same way). Installation is performed under your home directory.

Reserve some AMD GPU nodes with OAR

• Reserve a node with some AMD GPUs (see the Hardware page for the list of sites and clusters with GPUs).

flyon:

oarsub -I -l gpu=1 -t exotic -p "gpu_model like 'Radeon%'"

Please try to not reserve a single GPU on nodes with many GPUs (e.g. ≥ 4) if you only need to execute code on one GPU. For instance, using the neowise cluster is not very welcome for a user to use only one GPU at a time.

To reserve the full node (with all its GPUs):

flyon:

oarsub -I -l host=1 -t exotic -p "gpu_model like 'Radeon%'"

• Once connected to the node, check GPU presence:

neowise:

rocm-smi

======================= ROCm System Management Interface =======================
================================= Concise Info =================================
GPU  Temp   AvgPwr  SCLK    MCLK    Fan   Perf  PwrCap  VRAM%  GPU%  
0    26.0c  19.0W   930Mhz  350Mhz  255%  auto  225.0W    0%   0%    
================================================================================
============================= End of ROCm SMI Log ==============================

PyTorch

	Note
	Conda packages are not currently available for ROCm, please use pip instead

• Go on PyTorch website to see the installation command that suits you.

For instance (as of December 2021), selecting “Stable”, “Linux”, “Pip”, “Python”, “ROCM 5.4.2” gives this command to execute:

neowise:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

• Check if PyTorch is correctly installed to works with GPU:

neowise:

python3 -c "import torch; print('Num GPUs Available:', torch.cuda.device_count())"

Num GPUs Available: 8

Tensorflow

	Note
	On AMD GPU, Tensorflow is only supported using Docker images.

• Enable docker on your node (--tmp option is used to use /tmp directory for docker storage)

neowise:

g5k-setup-docker --tmp

• Start ROCm's Tensorflow as explained in https://hub.docker.com/r/rocm/tensorflow/

neowise:

alias drun='sudo docker run -it --network=host --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size 16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /tmp/dockerx:/dockerx'

neowise:

drun rocm/tensorflow:latest

• From within the Docker container, check if Tensorflow is correctly installed to works with GPU:

neowise:

python3 -c "import tensorflow as tf; print('Num GPUs Available:', len(tf.config.experimental.list_physical_devices('GPU')))"

 Num GPUs Available: 8

Deep learning on ppc64 nodes

About the ppc64 architecture

Grid'5000 has an IBM cluster (drac) with a total of 48 GPUs.

This cluster is using a ppc64 architecture, which is much less common than the usual x86_64 (amd64) architecture. In particular, many deep learning frameworks are primarily targeted at x86_64 and may be hard to use on ppc64.

As a result, if you want to use this cluster for deep learning, you should be ready to invest more time to setup your experiments compared to the usual x86_64 clusters.

Options to install deep learning tools

We provide installation guides for three popular deep learning frameworks: PyTorch, TensorFlow and MXnet.

In general, there are several methods to install deep learning tools, each with advantages and disadvantages:

• modules: we provide pre-built software stacks for several deep learning tools: this is the easiest way to use them. If you need specific versions or build options, contact us.
• IBM PowerAI conda channel: IBM provides a Conda channel with deep learning tools built for ppc64. It is easy to install, but the provided tools versions are often quite out-of-date.
• pip packages: a few tools provide pip packages for ppc64, but this is rare: most pip packages are only available for x86_64
• Docker images: we support installing Docker#Nvidia-docker including support for GPU. You will need to run ppc64 docker images though.
• build from source: this is for advanced users

See below for details on how to install each tool.

Reserve ppc64 GPU nodes with OAR

• Reserve a ppc64 node with GPUs (see the Hardware page of drac cluster for details).

To reserve a full node for one hour:

fgrenoble:

oarsub -I -p cluster=drac -t exotic -l host=1,walltime=1:00

• Once connected to the node, check GPU presence and the available CUDA version:

drac:

$ nvidia-smi

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.197.02   Driver Version: 418.197.02   CUDA Version: 11.2     |
+-----------------------------------------------------------------------------+
(...)

	Note
	Nodes in the drac cluster come with a known-working Nvidia driver version in their default environment. If you install a more recent driver or deploy your own images, you may experience frequent system crashes with recent Nvidia drivers on Debian or Ubuntu. CentOS seems unaffected by the crashes. See nvidia developer forum for details.

IBM PowerAI conda channel

IBM PowerAI provides a Conda channel with dedicated packages compiled for ppc64le: https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/#/

• Load and activate conda

drac:

module load conda

• Install a package '<package>' from IBM PowerAI

	Note
	do not forget to create a dedicated environment before.

drac:

conda install -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/ <package>

PyTorch on ppc64

Load pytorch from modules

• Some packages in PowerAI might require older dependencies. For instance, the version of PyTorch is too old for Python 3.8 or Python 3.9, we must use Python 3.7:

drac:

conda create --name pytorch-ppc64-py37 python=3.7

drac:

conda activate pytorch-ppc64-py37

drac:

conda install -c https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/ pytorch python=3.7

We provide a pre-built version of pytorch, and we can provide more versions on request. It is the easiest way to use pytorch as there is nothing to install.

As of November 2021, we provide pytorch 1.7.1. To use it:

drac:

module load python py-torch

drac:

python3 --version

Python 3.7.9

drac:

python3 -c 'import torch; print(torch.cuda.is_available())'

True

Note that you need to use the version of Python from our Modules, because Pytorch is built against Python 3.7 and won't work with the version of Python available in Debian 11 (Python 3.9).

That's it: your pytorch projects should now work while the module is loaded.

If you want to load the module in a non-interactive job, see Modules#Using_modules_in_jobs

Install pytorch from IBM PowerAI

PowerAI 1.7.0 provides pytorch 1.3.1

• To install it, load conda and create a Python 3.7 environment:

drac:

module load conda"

drac:

conda create --name pytorch-ppc64-py37 python=3.7

drac:

conda activate pytorch-ppc64-py37

• Add PowerAI repository:

drac:

conda config --prepend channels https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/

• Install pytorch:

drac:

conda install pytorch

• It will take around 10 minutes to download and install. Test that it works:

drac:

python3 -c "import torch; print(torch.cuda.is_available())"

True

	Note
	If this doesn't work, make sure that you are using the correct Python interpreter provided by Conda, using e.g. which python3. In some cases, you might have to specify the interpreter as python3.7.

Tensorflow on ppc64

Install from conda via IBM PowerAI

• PowerAI 1.7.0 provides tensorflow 2.1.3. It is the same principle as PyTorch, prepare a conda environment with Python 3.7:

drac:

module load conda conda create --name tensorflow-ppc64-py37 conda activate tensorflow-ppc64-py37 conda config --prepend channels https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/

• Install Tensorflow with GPU support:

drac:

conda install tensorflow-gpu

• It will take around 10 minutes to download and install. Test that it works:

drac:

python3 -c "import tensorflow as tf; print('Num GPUs Available:', len(tf.config.experimental.list_physical_devices('GPU')))"

Num GPUs Available: 4

Install from pip

Tensorflow is not available in pip for ppc64. However, we can use a non-official pip package. It provides a reasonably recent version: tensorflow 2.3.2.

Unfortunately, as of November 2021, this unofficial package is not compatible with Python 3.9. It means that we have to use Python 3.7 through modules as a workaround.

• Start by loading Python 3.7:

drac:

module load python

drac:

python3 --version

Python 3.7.9

• Then create a virtualenv:

drac:

python3 -m venv ~/venv-py3-tensorflow

drac:

. ~/venv-py3-tensorflow/bin/activate

• Then install Tensorflow from the non-official pip wheel:

drac:

wget https://powerci.osuosl.org/job/TensorFlow2_PPC64LE_GPU_Release_Build/20/artifact/tensorflow_pkg/tensorflow-2.3.2-cp37-cp37m-linux_ppc64le.whl

drac:

pip install --upgrade pip setuptools

drac:

pip install ./tensorflow-2.3.2-cp37-cp37m-linux_ppc64le.whl

It takes around 5-10 minutes to install because some dependencies need to be compiled.

• At runtime, you will need cudnn. You can install it yourself, or we provide it as a module for convenience:

drac:

module load cudnn

• Test that it works:

drac:

python3 -c "import tensorflow as tf; print('Num GPUs Available:', len(tf.config.experimental.list_physical_devices('GPU')))"

 Num GPUs Available: 4

As before, if you want to load cudnn in a non-interactive job, see Modules#Using_modules_in_jobs

Build tensorflow from source

The last option is to build tensorflow from source yourself, which is useful if you need a specific version or specific features. This is for advanced users and we provide no support.

It has been reported to work with a CentOS docker container using https://github.com/tensorflow/build/tree/master/ppc64le_builds

See https://github.com/anji993/build/tree/anji993-patch-1/ppc64le_builds for build instructions on Grid'5000.

Nvidia-docker for ppc64

Installation

To easily install Nvidia-docker on a node, see Docker#Nvidia-docker.

Running ppc64le Docker images

You need to make sure you are running Docker images that are built for ppc64le.

Example sources of ppc64le images:

• https://hub.docker.com/r/ibmcom/tensorflow-ppc64le
• https://hub.docker.com/u/nvidia/
• https://hub.docker.com/u/ppc64le/
• https://hub.docker.com/search/?type=image&architecture=ppc64le

To test tensorflow with an image from IBM:

drac:

tensorflowtest="import tensorflow as tf; print('Num GPUs Available:', len(tf.config.experimental.list_physical_devices('GPU')))"

drac:

docker run -it --rm --gpus all ibmcom/tensorflow-ppc64le:latest-gpu-py3 python -c "$tensorflowtest"

2021-02-15 11:33:10.853846: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1686] Adding visible gpu devices: 0, 1, 2, 3
Num GPUs Available: 4