Tensorflow Guide

From Storrs HPC Wiki
Revision as of 08:06, 3 May 2021 by Dbs18001 (talk | contribs) (Anaconda route)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search
Tensorflow
Author Google Brain Team
Website https://www.tensorflow.org/
Source HistoryVersions
Category Machine Learning, Deep Neural Networks
Help documentation


Current Supported Versions in HPC

CPU Support Version Corresponding Python Module Dependence Released Date
v0.10.0rc0 python/2.7.6 - Jul 29, 2016
- - - -
GPU & CPU Support Version Corresponding Python Module Dependence Released Date
v0.12.1 python/2.7.6-gcc-unicode cuda, cudnn Dec 25, 2016
v1.2.0 python/3.6.1 cuda, cudnn Jun 16, 2017

Import Modules

TensorFlow v0.10.0rc0 (CPU Support Only)

module load python/2.7.6

TensorFlow v1.2.0 (Both GPU and CPU Support)

module load cuda/8.0.61 cudnn/6.0 sqlite/3.18.0 tcl/8.6.6.8606 python/3.6.1

Or:

module load gcc/5.4.0-alt cuda/8.0.61 cudnn/6.0 sqlite/3.18.0 tcl/8.6.6.8606 python/3.6.3-gcc540a

TensorFlow v0.12.1 (Both GPU and CPU Support)

module load cuda/8.0.61 cudnn/6.0 python/2.7.6-gcc-unicode

Notice:

  • 1: This version supports both CPU and GPU.
  • 2: You should see the following auto-output, which means the gpu dependence libraries have been loaded successfully.
 I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally
 I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally
 I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally
 I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcudnn.so.6 locally
 I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally
  • 3: You will get the error below if you run python in normal computing node. Please do not worry about it! This just mean the gpu feature is unavailable due to current node isn't GPU node. You can still use cpu for calculation without any problem.
 E tensorflow/stream_executor/cuda/cuda_driver.cc:509] failed call to cuInit: CUresult(-1)
 I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:145] kernel driver does not appear to be running on this host (cn02): /proc/driver/nvidia/version does not exist

Application Usage

Check Version Number

 $ python
 Python 2.7.6 (default, Apr 20 2016, 16:39:52) 
 [GCC 4.8.2] on linux2
 Type "help", "copyright", "credits" or "license" for more information.
 >>> import tensorflow as tf
 >>> print tf.__version__
 0.12.1
 >>>

Example: How to Build Computational Graph

$ python
Python 2.7.6 (default, Apr 20 2016, 16:39:52) 
[GCC 4.8.2] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> node1 = tf.constant(3.0, tf.float32)
>>> node2 = tf.constant(4.0)
>>> node3 = tf.add(node1, node2)
>>> sess = tf.Session()
>>> sess.run(node3)
7.0
>>> a = tf.placeholder(tf.float32)
>>> b = tf.placeholder(tf.float32)
>>> adder_node = a + b
>>> print(sess.run(adder_node, {a: 3, b:4.5}))
7.5
>>> print(sess.run(adder_node, {a: [1,3], b: [2, 4]}))
[ 3.  7.]
>>> add_and_triple = adder_node * 3
>>> print(sess.run(add_and_triple, {a: 3, b:4.5}))
22.5
>>>

Example: Logging Device placement (GPU Version Guide)

hpc-xin@cn02:~$ ssh gpu01
hpc-xin@gpu01:~$ module purge
hpc-xin@gpu01:~$ module load cuda/8.0.61 cudnn/6.0 python/2.7.6-gcc-unicode
hpc-xin@gpu01:~$ python
Python 2.7.6 (default, Apr 20 2016, 16:39:52) 
[GCC 4.8.2] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcudnn.so.6 locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally
>>> a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
>>> b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
>>> c = tf.matmul(a, b)
>>> sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla K40m, pci bus id: 0000:03:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:1) -> (device: 1, name: Tesla K40m, pci bus id: 0000:82:00.0)
Device mapping:
/job:localhost/replica:0/task:0/gpu:0 -> device: 0, name: Tesla K40m, pci bus id: 0000:03:00.0
/job:localhost/replica:0/task:0/gpu:1 -> device: 1, name: Tesla K40m, pci bus id: 0000:82:00.0
I tensorflow/core/common_runtime/direct_session.cc:255] Device mapping:
/job:localhost/replica:0/task:0/gpu:0 -> device: 0, name: Tesla K40m, pci bus id: 0000:03:00.0
/job:localhost/replica:0/task:0/gpu:1 -> device: 1, name: Tesla K40m, pci bus id: 0000:82:00.0
>>> print(sess.run(c))
MatMul: (MatMul): /job:localhost/replica:0/task:0/gpu:0
I tensorflow/core/common_runtime/simple_placer.cc:827] MatMul: (MatMul)/job:localhost/replica:0/task:0/gpu:0
b: (Const): /job:localhost/replica:0/task:0/gpu:0
I tensorflow/core/common_runtime/simple_placer.cc:827] b: (Const)/job:localhost/replica:0/task:0/gpu:0
a: (Const): /job:localhost/replica:0/task:0/gpu:0
I tensorflow/core/common_runtime/simple_placer.cc:827] a: (Const)/job:localhost/replica:0/task:0/gpu:0
Const: (Const): /job:localhost/replica:0/task:0/cpu:0
I tensorflow/core/common_runtime/simple_placer.cc:827] Const: (Const)/job:localhost/replica:0/task:0/cpu:0
[[ 22.  28.]
 [ 49.  64.]]

You can run this code by tensorflow 1.2.0 in python 3.6.1 as well. You will get the same result but a little bit different intermediate automatic output.

More Examples

https://www.tensorflow.org/tutorials/

Python Packages depend on Tensorflow

Keras

As an example, here we will:

  • interactively run the MNIST example code from Keras: https://keras.io/examples/mnist_cnn/ - save the code into a file called mnist_cnn.py
  • Run with CPU only, then with a K40 GPU and finally with a V100 model GPU. The epoch processing speeds are:
  • CPU = ~5 minutes
  • GPU K40 = 10 seconds
  • GPU v100 = 3 seconds
srun -p gpu --gres gpu:1 --pty bash 
# srun: job 2886234 queued and waiting for resources
# srun: job 2886234 has been allocated resources
module purge
module load cuda/8.0.61 cudnn/6.0 tcl/8.6.6.8606 sqlite/3.18.0 python/3.6.1
which python
# Setting the empty CUDA_VISIBLE_DEVICES environmental variable below hides the GPU from TensorFlow so that we can run in CPU only mode.
CUDA_VISIBLE_DEVICES="" python mnist_cnn.py

For the GPU v100:

srun -p gpu_v100 --gres gpu:1 --pty bash
module purge
module load cuda/8.0.61 cudnn/6.0 tcl/8.6.6.8606 sqlite/3.18.0 python/3.6.1
which python
python mnist_cnn.py

Anaconda route

module purge
module load bazel/3.1.0 cuda/10.1 cudnn/7.6.5 gcc/8.4.0-rhel7 anaconda/5.1.0

Build an environment and then activate it! (Tensorflow GPU Version)

conda create -n tf_env -c anaconda -c conda-forge/label/gcc7 -c conda-forge/label/cf201901 python tensorflow-gpu=2.4.1 numpy pandas matplotlib seaborn scikit-learn scipy keras IPython jupyterlab tqdm pillow librosa pysoundfile pydub
source activate tf_env


Alternative (We can speed up solving if we specify all the versions.) (Tensorflow GPU Version)

conda create -n tf_env -c anaconda -c conda-forge/label/gcc7 -c conda-forge/label/cf201901 python=3.7.9 tensorflow-gpu=2.4.1 numpy=1.19.1 pandas=1.1.3 matplotlib=3.3.1 seaborn=0.11.0 scikit-learn=0.23.2 scipy=1.6.2 keras=2.4.3 ipython=7.18.1 jupyterlab=2.2.6 tqdm=4.50.2 pillow=8.0.0 librosa=0.6.2 pysoundfile=0.10.2 pydub=0.23.0
source activate tf_env

If verifying takes a long time and it migt help, it is possible to force skip this by running this before the conda create command.

conda config --set safety_checks disabled