Detection of Roles of Variables based on Deep Learning Technologies

A Labelling system & neural model designs

Recommended requirement

• Python 3.7.5
• TensorFlow 2
• numpy
• matplotlib
• sklearn

Labelling steps

1. Do Labelling in the source code, under folder selected-data/

2. Run the Labeller python main.py

3. dataset.npz is generated, under folder out-dataset/

Labelling

Here is the label id reference (Not all of them are used)

id	role
1	Fixed value
2	Stepper
3	Gatherer
4	Most-recent holder
5	Most-wanted holder
6	Follower
7	Temporary
8	One-way flag
9	-
10	-

Variable trace

To label a variable, replace the initial value to a construtor Intercept<int>(int initial_value, int label id) like the follow example.

for(int stepper = 0; stepper < 10; stepper++){
  //some operation
}

for(Intercept<int> stepper(0, 2); stepper < 10; stepper++){
  //some operation
}

Where Intercept<int> stepper(0, 2) is equal to Intercept<int> stepper = Intercept<int>(0, 2)

Some example can be found in demo-Intercept.cpp (see commented part) , running this program is easier to understand the class

State trace

Assuming you know about variable trace part

State trace is eventually putting variable traces into an array, and change the way to access the variable.
Examples can be found in demo-Intercept.cpp.

/* state trace labelling demo */
// labelling 2 variables, [0]: initial value = 1, role = 2 (stepper), [1]: initial value = 0, role = 3 (gatherer)
Intercept<int> intercepts[2] = {Intercept<int>(1, 2), Intercept<int>(0, 3)};
StateIntercept state(intercepts, 2);

// access variable using the class StateIntercept
state[0] = 10;

for (state[0]; state[0] > 0; state[0]--)
{
    state[1] += state[0];
}

Use the dataset

A class called DatasetLoader is recommended to access dataset.npz file easily, like this:

# assuming DatasetLoader.py is under the same folder
import DatasetLoader as loader

x, y, lengths, lengthMax = loader.variableTrace.load()
x, y, lengths, lengthMax, exeNames, roleInStates = loader.stateTrace.load()
x, y, lengths, lengthMax, exeNames, roleInStates = loader.stateTrace.load(model='2b')
(x1, x2), y, lengths, lengthMax, exeNames, roleInStates = loader.stateTrace.load(model='3')
(x1, x2), y, lengths, lengthMax, exeNames, roleInStates = loader.stateTrace.r5.load(model='3')

# predictions sample code
(x1, x2), y, lengths, lengthsMax, exeNames, roleInStates = loader.stateTrace.r5.prediction.load(model='3')
predictions = model.predict([x1,x2])
for i in range(120):
    print("--- Should be:", y[i])
    print("p1 = ",np.argmax(predictions[i],axis=-1), '\n' ,predictions[i])

See simple_rnn_model3.py for the latest dataset demo

See simple_autoencoder.py, simple_classifier.py, simple_demo1.py

The above program is built for variable trace, not sure if they work with state trace.

Model designs

Models' source code mentioned in our design

• rnn-model1.py
• rnn-model2a.py
• rnn-model2b.py
• rnn-model3a.py
• rnn-model3b.py
• rnn-model3c.py
• rnn-model3d.py