Titanic

Training a neural network on the Titanic dataset.

The dataset can be downloaded from https://www.kaggle.com/c/titanic. Since nnlib does not support feature engineering, the dataset needs to be prepared using the following Python script:

import numpy as np
import pandas as pd
import sys
from sklearn.preprocessing import MinMaxScaler
 
 
def scale(X, column):
    scaler = MinMaxScaler()
    transformed = scaler.fit_transform(X[:, column].reshape(-1, 1))
    X[:, column] = transformed[:, 0]
 
 
def main(argv):
    if len(argv) < 2:
        print('Filepath to the train file needs to be specified')
        return
 
    filepath = argv[1]
    print(f'Processing {filepath}')
 
    dataset = pd.read_csv(filepath)
    dataset['Age'].fillna(dataset['Age'].mean(), inplace=True)
    dataset = pd.get_dummies(dataset, prefix=['Sex'], columns=['Sex'], drop_first=True)
    dataset = dataset.drop(['PassengerId', 'Pclass', 'Name', 'Ticket', 'Cabin', 'Embarked'], axis=1)
    print(dataset.head())
 
    X = dataset.iloc[:, 1:].to_numpy()
    y = dataset.iloc[:, 0].to_numpy()
    print(X)
 
    scale(X, 0)  # Scale age
    scale(X, 1)  # Scale SibSp
    scale(X, 2)  # Scale Parch
    scale(X, 3)  # Scale Fare
    print(X)
 
    dfX = pd.DataFrame(X)
    dfy = pd.DataFrame(y)
    dfX['5'] = dfy  # Add the targets as the last column
    dfX.to_csv('./out/dataset.csv', header=None, index=False)
 
 
if __name__ == '__main__':
    main(sys.argv)
 

The Python script expects one parameter, which is the full path to the Titanic dataset. The dataset prepared for nnlib will then be generated in the ./out directory.

The main.cpp file expects one argument which is the absolute path to the file prepared by the Python script.

 
#include <iostream>
#include <nnlib/network.h>
#include <nnlib/read.h>
#include <nnlib/verify.cuh>
#include <nnlib/onehot_encode.h>
#include <chrono>
 
int main(int argc, char** argv) {
    if (argc < 2) {
        std::cout << "Dataset file path was not specified." << std::endl;
        return 1;
    }
 
    showCudaInfo();
 
    sTensor dataset = readCSV(argv[1], ",", 4);
    sTensor X = std::make_shared<Tensor>(dataset->shape[0], dataset->shape[1] - 1);
    sTensor y = std::make_shared<Tensor>(dataset->shape[0], 1);
 
    for (int i = 0; i < dataset->shape[0]; i++) {
        y->data[i] = dataset->data[i * dataset->shape[1] + dataset->shape[1] - 1];
        for (int j = 0; j < dataset->shape[1] - 1; j++) {
            X->data[i * X->shape[1] + j] = dataset->data[i * dataset->shape[1] + j];
        }
    }
 
    std::cout << y << std::endl;
 
    Network network = Network(X->shape[1]);
    network.add(64);
    network.add(y->shape[1], "sigmoid");
 
    std::vector<Metric*> metrics = {new BinaryAccuracy(), new MeanSquaredError()};
    network.train(X, y, 100, 10,  0.01, new BinaryCrossEntropy(), metrics);
 
    return 0;
}

The project can be built with the following CMake script. This script requires CMAKE_PREFIX_PATH to be set to the install directory of nnlib.

cmake_minimum_required(VERSION 3.20)
include(CheckLanguage)
 
project(titanic_nnlib LANGUAGES CXX)
set(CMAKE_CXX_STANDARD 14)
 
find_package(nnlib CONFIG REQUIRED)
 
check_language(CUDA)
if (CMAKE_CUDA_COMPILER)
    enable_language(CUDA)
    set(CMAKE_CUDA_STANDARD 14)
 
    include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
 
    # Add flag to specify the architecture of the GPU (compute capability)
    if (NOT DEFINED CUDA_ARCHITECTURES)
        set_target_properties(nnlib PROPERTIES CUDA_ARCHITECTURES "50")
    endif()
 
    set_target_properties(nnlib PROPERTIES
            CUDA_SEPARABLE_COMPILATION ON
            CUDA_RESOLVE_DEVICE_SYMBOLS ON)
endif()
 
add_executable(${PROJECT_NAME} main.cpp)
 
target_link_libraries(${PROJECT_NAME} PRIVATE nnlib)