U-Sharma/NeuralScaleID

NeuralScaleID

We implement teacher/student training with a variety of network sizes and shapes. We also test performance of CNNs on image datasets. The reader is strongly encouraged to refer to the example.ipynb in the codebase. More details of the codebase can be found in Documentation.md.

For the sake of organization, the python files for individual experients have been placed together in the folder Experiments. To run an experiment file that imports NeuralNet or teacher, it has to be in the same directory as those two modules. The organization of files in folder Experiments with respect to the paper is:

• KL.py: Teacher/student with cross-entropy loss
• L2.py: Teacher/student with mean squared error loss
• GeneralizedLoss.py: Teacher/student with generalized loss
• XxY.py: Teacher/student with product data teacher
• CIFAR10.py, FMNIST.py, MNIST.py and SVHN.py: CNN on the respective image datasets

Description of Teacher/Student modules

Generating a random teacher with 4 inputs and 2 outputs (softmax), and 2 hidden layers of width 128:

import teacher
t = teacher.Teacher(architecture=[4,128,128,2],softmax=True)

Argument scale can be used to control the number of features. The inputs and scale are multiplied element wise before passing through the network. In the example below, the first two input inputs are multiplied by 1 and the last two multiplied by 0. Thus, this example has 2 input features.

import teacher
t = teacher.Teacher(architecture=[4,128,128,2],softmax=True,scale=[1,1,0,0])

The module NeuralNet can be used to generate a fully connected neural network.

import NeuralNet as student
model = student.Model(architecture=[4,32,32,32,2])

This model can be trained on the teacher

x_train = t.generate_input(batch_size=32*10000)
y_train = t.predict(x_train)

history = model.train(x_train,y_train)

We can evaluate the loss on the trained model

x_test = t.generate_input(batch_size=100)
y_test = t.predict(x_test)

test_loss = model.evaluate(x_test,y_test)

The teacher/student experiments have been done by first generating a database from teacher, and then loading the database in order to train the student (refer example.ipynb). In order to run those experiments, a database needs to be generated. Refer to the individual experiment python files for details of database.