TensorFlow

TensorFlow Software

TensorFlow is an end-to-end open source platform for machine learning. It has a comprehensive, flexible ecosystem of tools, libraries and community resources that lets researchers push the state-of-the-art in ML and developers easily build and deploy ML powered applications. 

TensorFlow Versions

As TensorFlow leverages GPU, it is available in the 'allnodes' optimization tree.  Use the 'module avail' command to see the current list of versions (Output from 4 Nov 22):

----------------------- /usr/local/easybuild_allnodes/modules/all -----------------------

   TensorFlow/2.6.0-foss-2021a-CUDA-11.3.1    TensorFlow/2.7.1-foss-2021b-CUDA-11.4.1 (D)

Running MNIST Example in HPC

Copy the python script "mnist.py" from /usr/local/doc/TENSORFLOW/ to your home directory. It is adapted from TensorFlow 2 quickstart for beginners.

cp  /usr/local/doc/TENSORFLOW/mnist.py  .

Request an appropriate GPU node. In this sample,  gpu2v100 is used

srun -p gpu -C gpu2v100 --gres=gpu:1 --mem=8gb --pty bash

Load  Python module if you have not already loaded it along with appropriate CUDA and cuDNN modules.

module load TensorFlow/2.7.1-foss-2021b-CUDA-11.4.1

Run Python script:

python mnist.py

output:

9 MB memory:  -> device: 0, name: Tesla V100-PCIE-32GB, pci bus id: 0000:3b:00.0, compute capability: 7.0

Epoch 1/5

1875/1875 [==============================] - 3s 1ms/step - loss: 0.2964 - accuracy: 0.9145

Epoch 2/5

1875/1875 [==============================] - 2s 1ms/step - loss: 0.1416 - accuracy: 0.9580

Epoch 3/5

1875/1875 [==============================] - 2s 1ms/step - loss: 0.1071 - accuracy: 0.9675

Epoch 4/5

1875/1875 [==============================] - 2s 1ms/step - loss: 0.0870 - accuracy: 0.9731

Epoch 5/5

1875/1875 [==============================] - 2s 1ms/step - loss: 0.0742 - accuracy: 0.9769

313/313 - 0s - loss: 0.0733 - accuracy: 0.9777 - 413ms/epoch - 1ms/step

For batch submission, follow the instruction at HPC Batch & Interactive Job submission, and be aware of the specifics of Working with Modules on Pioneer

Tensorflow from a singularity Container

Please visit our HPC Guide to Singularity.

Tensorflow installation with pip

Tensorflow can also be installed using python virtual environment and pip install.

The instruction is given on this page: https://www.tensorflow.org/install/pip

If you want to run tensorflow with GPU, you would need to install it on a GPU node 

srun -p gpu --gres=gpu:1 --mem=16gb --pty bash

nvidia-smi

module load Python

python3 -m venv tf 

source tf/bin/activate

pip install --upgrade pip

# For GPU users

pip install tensorflow[and-cuda]

python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

The last part of the output:

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]