In this chapter, we will introduce an example of computing using WHEEL, “MNIST data analysis workflow using TensorFlow.”

1. Overview

Computation Overview

In this case, we use TensorFlow on Fugaku to create and infer a handwritten digit image learning model using MNIST data.

This tutorial assumes that you are running on Mt. Fugaku. If you are running in a different environment, change the script settings accordingly.

Also, MNIST data is an image data set consisting of 60,000 handwritten digit images and 10,000 test images, and is famous as a very useful image data set for learning neural networks. The data set contains the following structure:

MNIST data
　├ Learning data (60,000 pieces)
　│ 　├ image data
　│ 　└ Label data (correct answer data)
　│
　└ Validation data (10,000 pieces)
　　 　├ image data
　　 　└ Label data (correct answer data)

For each image data, label data which is correct data is given one by one.

Workflow Overview

This tutorial uses two task components:

train component
Create a machine learning model.
estimate component
Validate the model created by train component.

img

2. About Neural Network (NN)

This section describes the NN method used to analyze MNIST data. NN stands for neural network, which is a mathematical model of artificial neurons representing nerve cells (neurons) and their connections, or networks, in the human brain. It consists of an input layer, an output layer, and a hidden layer, with a “weight” between the layers that indicates the strength of the connections between the neurons.

img

Features are extracted from the image, and the weights of nodes are updated and learned. In this case, the MNIST data set is used, so the input layer is 784 nodes and the output layer is 10 nodes in order to obtain an output of one of 0~9 digits from an input of 28*28 pixels.

3. Workflow Creation

Create a new project and add two task components. Each component is named train and estimate.

Configuring the train component

Create a script for the train component. Create a new file named run.sh in the train component and fill in the following:

#PJM --rsc-list "elapse=01:00:00"
#PJM -L "node=1"
#PJM -x PJM_LLIO_GFSCACHE=/vol0004


export PATH=/home/apps/oss/TensorFlow-2.2.0/bin:$PATH
export LD_LIBRARY_PATH=/home/apps/oss/TensorFlow-2.2.0/lib:$LD_LIBRARY_PATH

cat <<EOF >kerasMnist.py
from tensorflow import keras
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, InputLayer
from tensorflow.keras.optimizers import RMSprop

print('download data')
---
title: load mnist data
---
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print('download done')

---
title: set input data
---
x_train = x_train.reshape(60000, 784)
x_test = x_test.reshape(10000, 784)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
y_train = keras.utils.to_categorical(y_train, 10)
y_test = keras.utils.to_categorical(y_test, 10)

---
title: create model
---
model = Sequential()
model.add(InputLayer(input_shape=(784,)))
model.add(Dense(10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])


---
title: train model
---
print('start learning')
epochs = 20
batch_size = 128
history = model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1, validation_data=(x_test, y_test))

---
title: save weights
---
print('save weights')
model.save_weights('param.hdf5')
EOF

python kerasMnist.py > learn.log

Open the component properties window and set the following five items.

  • script: run.sh
  • host: fugaku
  • use job scheduler: Enabled
  • output files: param.hdf5
  • include: learn.log

estimate component settings

Create a script for the estimate component. Create a new file named run.sh in the estimate component and fill in the following:

#PJM --rsc-list "elapse=01:00:00"
#PJM -L "node=1"
#PJM -x PJM_LLIO_GFSCACHE=/vol0004


export PATH=/home/apps/oss/TensorFlow-2.2.0/bin:$PATH
export LD_LIBRARY_PATH=/home/apps/oss/TensorFlow-2.2.0/lib:$LD_LIBRARY_PATH

cat <<EOF >kerasMnist.py
from tensorflow import keras
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, InputLayer
from tensorflow.keras.optimizers import RMSprop

print('download data')
---
title: load mnist data
---
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print('download done')

---
title: set input data
---
x_test = x_test.reshape(10000, 784)
x_test = x_test.astype('float32')
x_test /= 255
y_test = keras.utils.to_categorical(y_test, 10)

---
title: create model
---
model = Sequential()
model.add(InputLayer(input_shape=(784,)))
model.add(Dense(10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

---
title: load weights
---
print('loadweights')
model.load_weights('param.hdf5')

---
title: eval model
---
print('eval')
score = model.evaluate(x_test, y_test, verbose=1)
print()
print('Test loss:', score[0])
print('Test accuracy:', score[1])
EOF

python kerasMnist.py > eval.log

Open the component properties window and set the following four items.

  • script: run.sh
  • host: fugaku
  • use job scheduler: Enabled
  • include: eval.log

Setting File Dependencies

Finally, if you drop ▶ in param.hdf5 set to output files in train and connect it to the estimate component, the workflow creation is complete.

img

4. Checking the execution results

Review the analysis results. After execution is complete, you can see the following results from the log.kerasMNIST and log.kerasMNISTEval files created in the remote task.

・
・
・
('Test loss:', 0.27075190788507464)
('Test accuracy:', 0.927)

The meanings of the terms are as follows:

Property name Setting value
Test loss Loss value when learning data is supplied. The smaller the number, the more correct the result.
Test accuracy Percentage of correct data for validation. A value closer to 1 indicates higher accuracy in the validation data.

The above results confirm that accuracy is 0.927, that is, a learning model with 92.7% accuracy in the validation data was created.

That’s all for an example MNIST data analysis workflow using TensorFlow.

Return to Practical Tutorial

Updated: