Andrew Pandji

Exploring spatial AI

Training a Neural Network

Andrew Pandji

2 min read ·

Training a Neural Network

Training a Shape Classifier

  • Shape Classifier that distinguishes between circles and rectangles using 28x28 thumbnails. Similar method for more advanced types of classifiers.
  • Process:
    • Data Gathering
    • Data Cleaning & Preprocessing
      • Removing dogs from scraped set of cat pictures
    • Data Augmentation (For training data)
      • Geometric transformation
      • Color/Brightness adjustment
      • Synthetic Noise addition
    • Dataset Splitting
      • Training Set (70-80%)
      • Validation Set (10-15%)
      • Test Set (10-15%)
    • Final Datasets (Training, Validation, Test)
    • Deep Learning Model Training -> this is iterative
    • Model Evaluation & Testing
  • When developing a deep learning model - use 80% for training and 20% for validation.
  • Labeling & Splitting Training Data
    • predict
      • a random set of images
    • train
      • CIRCLE
        • the image itself, and the label (/CIRCLE, the name itself doesn’t matter circle_0000.png)
      • RECTANGLE
  • Natural vs Synthetic Data
    • Natural - is data out in the world, gathered through observation
    • Synthetic - generated in various ways, can be using AI or just simple scripts (Processing, etc.)
  • Getting Started
import os // just bringing in the os
import numpy as np // numpy is common for vector and matrix math
import pandas as pd // pandas is tabular data
import cv2 // standard machine vision library, image manipulation
import tensorflow as tf // deep learning and AI
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dropout, Dense // simpler interface for deep learning / AI
from tensorflow.keras.models import Sequential
from sklearn.model_selection import train_test_split // scikit Classical Machine learning
from sklearn.preprocessing import LabelEncoder
  • Machine Learning vs. AI
    • Machine Learning is glorified statistics - clustering, dimension reduction, finding correlations in data, light pattern recognition
    • Whereas “deep learning” is machine learning x 1000, using the same techniques but stacking it sequentially so you can do deeper analysis
  • To initiate the process
def start(self, retrain=True); // Train a model or load existing

def prepare_data(self); // Map labels to integers, split training and testing data

def train_model(self); // Train the model

def predict(self, path); // Classify an image, with trained model

def predict_all(self); // Classify all images in a folder

def process_prediction(self, image_path); // executes prediction and finds the label

def image_to_input(self, image_path); // resize and grayscale images to prep into same format

def read_images(self); //

def images_to_csv(self, image_path, label) // transform an image to numberic matrix
  • An Epoch is the retraining to he model through every element of the data set
  • Training creates an .h5 file
    • For javascript -> need an onyx file
    • For Swift?
  • scikit-learn, opencv-python,
  • Processing can be very useful for generating and transforming image based datasets
Get in touch