TensorFlow-Keras
This section describes how to install and run TensorFlow and Keras, and how to install TensorFlow, Keras and Horovod to perform distributed learning.
Running TensorFlow-Keras on a single node
Precondition
- Replace
grpname
with your own ABCI group. - The Python virtual environment should be created in the home or group area so that it can be referenced by interactive nodes and each compute node.
- The sample program should be saved in the home or group area so that it can be referenced by interactive nodes and each compute node.
Installation
Here are the steps to create a Python virtual environment and install TensorFlow and Keras into the Python virtual environment.
[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5
[username@g0001 ~]$ python3 -m venv ~/venv/tensorflow-keras
[username@g0001 ~]$ source ~/venv/tensorflow-keras/bin/activate
(tensorflow-keras) [username@g0001 ~]$ pip3 install --upgrade pip setuptools
(tensorflow-keras) [username@g0001 ~]$ pip3 install tensorflow-gpu==1.15 keras
With the installation, you can use TensorFlow and Keras next time you want to use it by simply loading the module and activating the Python virtual environment, as follows.
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5
[username@g0001 ~]$ source ~/venv/tensorflow-keras/bin/activate
Execution
The following shows how to execute the TensorFlow sample program mnist_cnn.py
in the case of an interactive job and a batch job.
Run as an interactive job
[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5
[username@g0001 ~]$ source ~/venv/tensorflow-keras/bin/activate
(tensorflow-keras) [username@g0001 ~]$ git clone https://github.com/keras-team/keras.git
(tensorflow-keras) [username@g0001 ~]$ python3 keras/examples/mnist_cnn.py
Run as a batch job
Save the following job script as a run.sh
file.
1 2 3 4 5 6 7 8 9 10 11 12 |
|
Submit a saved job script run.sh
as a batch job with the qsub command.
[username@es1 ~]$ qsub -g grpname run.sh
Your job 1234567 ('run.sh') has been submitted
Running TensorFlow on multiple nodes
Precondition
- Replace
grpname
with your own ABCI group. - The Python virtual environment should be created in the home or group area so that it can be referenced by interactive nodes and each compute node.
- The sample program should be saved in the home or group area so that it can be referenced by interactive nodes and each compute node.
Installation
Here are the steps to create a Python virtual environment and install TensorFlow, Keras and Horovod into the Python virtual environment.
[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5 nccl/2.5/2.5.6-1 openmpi/2.1.6 gcc/7.4.0
[username@g0001 ~]$ python3 -m venv ~/venv/tensorflow-keras+horovod
[username@g0001 ~]$ source ~/venv/tensorflow-keras+horovod/bin/activate
(tensorflow-keras+horovod) [username@g0001 ~]$ pip3 install --upgrade pip setuptools
(tensorflow-keras+horovod) [username@g0001 ~]$ pip3 install tensorflow-gpu==1.15 keras
(tensorflow-keras+horovod) [username@g0001 ~]$ HOROVOD_WITH_TENSORFLOW=1 HOROVOD_GPU_OPERATIONS=NCCL HOROVOD_NCCL_HOME=$NCCL_HOME pip3 install --no-cache-dir horovod
With the installation, you can use TensorFlow, Keras and Horovod next time you want to use it by simply loading the module and activating the Python virtual environment, as follows.
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5 nccl/2.5/2.5.6-1 openmpi/2.1.6 gcc/7.4.0
[username@g0001 ~]$ source ~/venv/tensorflow-keras+horovod/bin/activate
Execution
The following shows how to execute a sample program keras_mnist.py
of TensorFlow with Horovod for distributed learning.
Run as an interactive job
In this example, using 4 GPUs in an interactive node for distributed learning.
[username@es1 ~]$ qrsh -g grpname -l rt_G.large=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.6/3.6.5 cuda/10.0/10.0.130.1 cudnn/7.6/7.6.5 nccl/2.5/2.5.6-1 openmpi/2.1.6 gcc/7.4.0
[username@g0001 ~]$ source ~/venv/tensorflow-keras+horovod/bin/activate
[username@g0001 ~]$ git clone -b v0.20.0 https://github.com/horovod/horovod.git
[username@g0001 ~]$ mpirun -np 4 -map-by ppr:4:node python3 horovod/examples/keras_mnist.py
Run as a batch job
In this example, a total of 8 GPUs are used for distributed learning. 2 compute nodes are used, with 4 GPUs per compute node.
Save the following job script as a run.sh
file.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
|
Submit a saved job script run.sh
as a batch job with the qsub command.
[username@es1 ~]$ qsub -g grpname run.sh
Your job 1234567 ('run.sh') has been submitted