Experiments with Convolutional Neural Network in TensorFlow

TensorFlow Project
Experiments with Convolutional Neural Network in TensorFlow
Background
Convolutional neural network is a powerful learning mechanism that was applied successfully to many computer vision tasks. In this project you are asked to run experiments using tensorflow. The experiments are
closely related to an example given in the TensorFlow tutorial.
The goal is to train a CNN to recognize images of digits from the MNIST dataset. This dataset consists of
gray level images of size 28x28. There is a standard partitioning of the dataset into training and testing.
The training has 60,000 examples and the testing has 10,000 examples. Standard CNN models achieves over
99% accuracy on this dataset.
In this project you are asked to solve a similar problem of creating a classifier for the MNIST data. However, the training data in our case has only 6,000 images, and each image is shrunk to size 7x7. Specifically,
your program must include the following:
1. Your program must set the random seeds of python and tensorflow to 1 to make sure that your results
are reproducible.
2. The first layer in the network must be a 4 × 4 maxpooling layer. This effectively shrinks the images
from 28x28 to 7x7.
3. Your program will be tested by training on a fraction of 0.1 of the standard training set. The testing
data will be the entire standard testing set.
4. The training and testing in you program should not take more than 6 minutes.
Provided programs and data
1. The original 60,000 training examples are in the folder big.
2. A random subset of 6,000 the training examples is in the folder small.
3. An example program that satisfies the project requirements is available as proj-tf1.py.
4. A program that can extract a random fraction from the training data is available as fraction xy.py.
What you need to do
Design a network to solve this problem. You can use all the functionality of tensorflow, not only the parts
that were described in class.
Grading
We will generate a random set of 6,000 training examples by running the program fraction xy.py with a seed
that is kept secret. If, for example, the seed is 7, the program is run as follows:
python3 fraction\_xy.py big/x\_train.csv big/y\_train.csv 0.1 7
Your grade will be based on the accuracy of your model trained with the generated examples and tested
on the standard testing data.
What you need to submit
1. Source code of the python script.
2. Documentation describing your network, and the results of experiments/accuracy that your program
achieves on the provided data.
1