Neural Network Project

The goal here is to generate additional data samples from a structured dataset using a Generative Adversarial Network. See our poster here.

Replication instructions

Set up

Begin by cloning the repo and installing the packages

git clone https://github.com/despresj/ece_colab.git
python3 -m venv ece884_project_enviroment
source ece884_project_enviroment/bin/activate
pip install -r requirements.txt

After that load some libraries.

import os
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler

Then import the functions created for this project. Read the code here

from tools.train import build_network, train_gan # functions from project

import and scale the data. Note that any tabular data set with numeric columns will (hopefully) work.

df = pd.read_csv("data_clean/taxi_sample.csv") # data is a small subset of the full nyc taxi data
# obtain the full data for 2016 here
data_columns = df.columns
scaler = MinMaxScaler().fit(df)
df = scaler.transform(df)

Train and generate data. The project was done in Google Colab, however this is designed to produce output locally. For the best results use the full dataset.

for _ in range(5000):
    # these GANS often converge to one point, these randomized hyper parameters
    # prevent it getting stuck on one solution. Also, these are appropriate for 
    # this taxi data another  dataset would require different randomized archatecture
    neurons = np.random.randint(8, 25)
    hidden = np.random.randint(5, 10)
    noise_n = np.random.randint(125, 150)
    epochs = np.random.randint(75, 100)
    learna = np.random.exponential(1e-2)
    learnb = np.random.exponential(1e-2)

    generator = build_network(output_dim=df.shape[1], n_hidden=hidden, n_neurons=neurons, learning_rate=learna)
    discriminator = build_network(output_dim=1, n_hidden=hidden, n_neurons=neurons, learning_rate=learnb) 
    
    gen_data = train_gan(generator, discriminator, df, n_epochs=epochs, n_noise=noise_n)
    output_path = "generated_data.csv"
    generated_data = pd.DataFrame(scaler.inverse_transform(gen_data), columns=data_columns) 
    generated_data.to_csv(output_path, mode='a', header=not os.path.exists(output_path), index=False)

Basic Structure of a GAN[2]

See the majority of the data generated are plausable.

Although plausable, these require significant curiation hence the cavilar approach to the neural network archatectr.