CSCI-566 Assignment 2

The objectives of this assignment

Implement the forward and backward passes as well as the neural network training procedure for Recurrent Neural Networks (RNNs).
Learn the basic concepts of language modeling and how to apply RNNs.
Implement popular a generative model, Generative Adversarial Networks (GANs) using TensorFlow.

Work on the assignment

Please first clone or download as .zip file of this repository.

Working on the assignment in a virtual environment is highly encouraged. In this assignment, please use Python 3.5 (or 3.6). You will need to make sure that your virtualenv setup is of the correct version of python.

Please see below for executing a virtual environment.

cd CSCI566-Assignment2
pip3 install virtualenv # If you didn't install it
virtualenv -p $(which python3) ./venv_cs566_hw2
source ./venv_cs566_hw2/bin/activate

# Install dependencies
pip3 install -r requirements.txt

# install tensorflow (cpu version, recommended)
pip3 install tensorflow=='1.14.0'

# install tensorflow (gpu version)
# run this command only if your device supports gpu running
pip3 install tensorflow-gpu=='1.14.0'

# Work on the assignment

# Deactivate the virtual environment when you are done
deactivate

Work with IPython Notebook

To start working on the assignment, simply run the following command to start an ipython kernel.

# add your virtual environment to jupyter notebook
python -m ipykernel install --user --name=venv_cs566_hw2

# port is only needed if you want to work on more than one notebooks
jupyter notebook --port=<your_port>

and then work on each problem with their corresponding .ipynb notebooks. Check the python environment you are using on the top right corner. If the name of environment doesn't match, change it to your virtual environment in "Kernel>Change kernel".

Problems

In each of the notebook file, we indicate TODO or Your Code for you to fill in with your implementation. Majority of implementations will also be required under lib with specified tags.

Problem 1: RNNs for Language Modeling (60 points)

The IPython Notebook Problem_1.ipynb will walk you through implementing a recurrent neural network (RNN) from scratch.

Problem 2: Generative Adversarial Networks (40 points)

The IPython Notebook Problem_2.ipynb will help you through implementing a generative adversarial network (GAN).

PLEASE DO NOT CLEAR THE OUTPUT OF EACH CELL IN THE .ipynb FILES

Your outputs on the .ipynb files will be graded. We will not rerun the code. If the outputs are missing, that will be considered as if it is not attempted.

How to submit

Run the following command to zip all the necessary files for submitting your assignment.

sh collectSubmission.sh

This will create a file named assignment2.zip, please rename it with your usc student id (eg. 4916525888.zip), and submit this file through the Google form. Do NOT create your own .zip file, you might accidentally include non-necessary materials for grading. We will deduct points if you don't follow the above submission guideline.

Questions?

If you have any question or find a bug in this assignment (or even any suggestions), we are more than welcome to assist. Please take a look at the FAQ section below before posting a question.

Again, NO INDIVIDUAL EMAILS WILL BE RESPONDED.

PLEASE USE PIAZZA TO POST QUESTIONS (under folder assignment2).

FAQ

Can I reuse the virtualenv from Assignment 1?
You can reuse the vistual environment but maybe you need to install some missing packages using pip3 install -r requirements.txt.
Maybe simpler is to create a new virtualenv, we give instructions above.
My RNN in Problem 1 is better than the LSTM?
Try experimenting with the number of training epochs (LSTM may train slower in the beginning) and the training prediction horizon (benefits of LSTMs get more apparent on longer prediction problems).
Even then, it is possible that the training dataset is too simple to show large benefits of LSTMs.
When parsing the text data file for Problem 1 I get a charmap codec can't decode error.
This might be a platform dependent issue. In past years adding encoding="utf8" to the file open command helped in these cases.
What is the meta variable used for in Problem 1?
This variable is used to pass all values to the backward pass that are necessary to compute the gradients. You can use it as a dictionary to pass any desired values over to the backward function.
I experimented with the hyperparameters and tried many different combinations, which ones should I report?
The usual rule of thumb is to report results with the best hyperparameters you found.
Exception is the prediction horizon parameter T in Problem 1, please do not report results for T smaller than the default value.
The function set_seed() produces an error?
Make sure your tensorflow version is not below 1.12.0.
My reconstruction loss for Problem 2.2 is higher than 32?
You should achieve a reconstruction loss lower than 32 finally.

thotamohan/GAN-Implementation