04-Multi-GPU processing

This guide demonstrates the use of several GPUs for a classification task with a simple convolutional network, based on this tutorial.

Jupyter Notebook preprocessing

We are going to build a simple convolutional network with the well known MNIST dataset to recognize handwritten digits.

First, we are going to download the source code from the notebook.

$ conda activate artemisa-tuto
(artemisa-tuto) $ wget https://raw.githubusercontent.com/keras-team/keras-io/master/examples/vision/ipynb/mnist_convnet.ipynb -O simple_convnet.ipynb

Once the notebook is downloaded, we can transform it into a Python file with the nbconvert tool.

(artemisa-tuto) $ pip install notebook
(artemisa-tuto) $ jupyter nbconvert simple_convnet.ipynb --to python --no-prompt

generating a Python file that only contains the python cells.

Multi-GPU job submission

Subsequently, we have to prepare the .sh and .sub. files.

In the .sh file we include the instructions to activate the conda environment and execute the generated .py file.

.sh executable

#!/bin/bash
EXERCISE_ENVIRONMENT="artemisa-tuto"
eval "$(conda shell.bash hook)"
conda activate $EXERCISE_ENVIRONMENT
python3 simple_convnet.py

In the submit description file we specify the number of gpus wanted for the job with the command request_gpus = 2. In order to increase the job throughput of the Artemisa facility, multiGPU job usage is restricted, accepting 2 and 4 GPUs jobs. If a 8-GPU job is needed, please contact us at artemisa-evaluacion@ific.uv.es explaining your specific computation needs and time slot required.

Submit description file

universe = vanilla

executable              = simple_convnet.sh

log                     = condor_logs/simple_convnet.log
output                  = condor_logs/simple_convnet.$(Cluster).$(Process).out
error                   = condor_logs/simple_convnet.$(Cluster).$(Process).err

getenv = True

request_gpus = 2
queue

After giving execution permits to the .sh file, we can submit the job

(artemisa-tuto) $ chmod +x simple_convnet.sh
(artemisa-tuto) $ condor_submit simple_convnet.sub
Submitting job(s).
1 job(s) submitted to cluster 578870.

After the job is finished, the output can be found in condor_logs/*.out file, which shows the characteristics of the convolutional network: shape, train and test samples, model and layers description, epochs run and finally the evaluation of the model, which achieves a ~99% accuracy.

x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
Model: "sequential"
_________________________________________________________________
Layer (type)                Output Shape              Param #
=================================================================
conv2d (Conv2D)             (None, 26, 26, 32)        320

max_pooling2d (MaxPooling2  (None, 13, 13, 32)        0
D)

conv2d_1 (Conv2D)           (None, 11, 11, 64)        18496

max_pooling2d_1 (MaxPoolin  (None, 5, 5, 64)          0
g2D)

flatten (Flatten)           (None, 1600)              0

dropout (Dropout)           (None, 1600)              0

dense (Dense)               (None, 10)                16010

=================================================================
Total params: 34826 (136.04 KB)
Trainable params: 34826 (136.04 KB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
Epoch 1/15
422/422 [==============================] - 5s 5ms/step - loss: 0.3746 - accuracy: 0.8859 - val_loss: 0.0847 - val_accuracy: 0.9775
Epoch 2/15
422/422 [==============================] - 1s 3ms/step - loss: 0.1104 - accuracy: 0.9664 - val_loss: 0.0578 - val_accuracy: 0.9845
Epoch 3/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0853 - accuracy: 0.9736 - val_loss: 0.0495 - val_accuracy: 0.9858
Epoch 4/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0701 - accuracy: 0.9788 - val_loss: 0.0419 - val_accuracy: 0.9892
Epoch 5/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0639 - accuracy: 0.9803 - val_loss: 0.0347 - val_accuracy: 0.9893
Epoch 6/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0562 - accuracy: 0.9828 - val_loss: 0.0348 - val_accuracy: 0.9902
Epoch 7/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0508 - accuracy: 0.9841 - val_loss: 0.0341 - val_accuracy: 0.9890
Epoch 8/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0484 - accuracy: 0.9851 - val_loss: 0.0332 - val_accuracy: 0.9900
Epoch 9/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0443 - accuracy: 0.9864 - val_loss: 0.0299 - val_accuracy: 0.9922
Epoch 10/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0396 - accuracy: 0.9871 - val_loss: 0.0289 - val_accuracy: 0.9933
Epoch 11/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0402 - accuracy: 0.9871 - val_loss: 0.0282 - val_accuracy: 0.9922
Epoch 12/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0378 - accuracy: 0.9879 - val_loss: 0.0271 - val_accuracy: 0.9923
Epoch 13/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0354 - accuracy: 0.9884 - val_loss: 0.0296 - val_accuracy: 0.9920
Epoch 14/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0337 - accuracy: 0.9896 - val_loss: 0.0292 - val_accuracy: 0.9910
Epoch 15/15
422/422 [==============================] - 1s 2ms/step - loss: 0.0324 - accuracy: 0.9895 - val_loss: 0.0258 - val_accuracy: 0.9925
Test loss: 0.02214638702571392
Test accuracy: 0.9926999807357788

Summary

Recap

It is possible to gather the code cells inside a Jupyter Notebook and transform it into a .py file, enabling the submission to the WN.
We can specify the number of GPUs needed with request_gpus, accepting 1, 2 and 4 GPU jobs. If a 8 GPU job is needed, please send a mail to artemisa-evaluacion@ific.uv.es.