PyTorch Tutorial - Neural Networks & Deep Learning in Python

Explore NumPy, pandas, Matplotlib, and Seaborn for data science. Install them with Anaconda and use SciPy, Statsmodels, scikit-learn, and H2O for statistics, regression, machine learning, and basic deep learning.

Explore numpy basics for data science, including multi-dimensional arrays and ndarray operations, and master import conventions and core functions like np.array and np.zeros.

Learn to create NumPy arrays, from rank-one to multi-dimensional ndarrays, using np.array, inspect shapes, and build zeros, diagonals, and basic linear algebra tools for data science.

Explore numpy operations, including creating arrays with np.array, indexing and slicing with zero and negative indices, and multidimensional access. Learn to concatenate arrays and understand shape, copying, and practical exercises.

Explore numpy-based vector arithmetic, including elementwise addition, subtraction, scalar operations, hadamard product, and dot product on one-dimensional vectors.

Explore numpy matrix arithmetic with two-dimensional matrices, including elementwise addition, subtraction, and multiplication, and scalar operations. Use hadamard multiplication and the np.dot multiply approach, plus transpose and inverse capabilities.

Introduce tensors as a generalization of vectors and matrices, viewed as multi-dimensional arrays. Show how tensors relate to numpy arrays and mention potential gpu acceleration for faster computations.

Create numpy arrays and PyTorch tensors, and convert between them with numpy and torch functions. Generate 3x3 matrices with ones or random values, and perform numpy or tensor conversions.

Explore basic PyTorch tensor operations, including creating tensors with ones and random values, reshaping and flattening, and performing elementwise addition, subtraction, multiplication, division, and mean, paralleling numpy behavior.

Learn to read csv data in Python with pandas, handling comma separated values and custom separators like semicolons or tabs, loading rest2.csv and previewing data with head.

Read Excel data using pandas by loading the Boston one.xls file from your working directory, print sheet names, and load a chosen sheet into a dataframe to view head.

Explore basic data pre-processing with pandas by reading CSV files, inspecting data types, dropping columns, handling missing values, and imputing with mean, median, and mode on the Titanic dataset.

Explore the theory of linear regression, showing how y changes with x and extending to multiple regression, using least squares, intercept, slope, and R-squared.

Learn to implement linear regression in python using pandas dataframes, fitting simple and multiple regression with statsmodels and sklearn, interpreting r-squared, p-values, and coefficients on iris data.

Explore gradient descent for ordinary least squares linear regression from first principles, linking input features to target values via a mean squared error cost and iterative weight updates.

Demonstrate ordinary least squares linear regression from first principles on toy data with one predictor, using train-test split and gradient descent in TensorFlow to minimize mean squared error.

Demonstrate implementing ordinary least squares regression from first principles in PyTorch, using a toy dataset, MSE loss, SGD optimization, and 10,000 training epochs to fit and visualize predictions.

Import pandas and numpy, load real CSV data, convert to PyTorch tensors, and implement ordinary least squares regression with autograd, a linear model, MSE loss, and SGD optimization.

Explore generalized linear models to handle non-normal residuals using exponential family distributions and link functions, with practical examples in logistic and Poisson regression for count and binary data.

Explore logistic regression as a GLM for binary categorical outcomes, using the probability and logit formula, odds, and core assumptions.

Learn to implement logistic regression in PyTorch on the Mnist dataset, converting images to tensors, building a model, and training with cross-entropy loss and SGD to reach about 89% accuracy.

Present the biology-based foundations of artificial neural networks, cover activation functions and thresholds, backpropagation, perceptrons, multilayer perceptrons, deep networks, plus NLP with PCA on large datasets.

Learn PyTorch syntax by importing torch and torch.nn as nn, defining an nn.Module with init and forward, and implementing a simple tensor multiplication to illustrate a neural network layer.

Explore activation functions in neural networks, including sigmoid, tan hyperbolic, Relu, and softmax, and learn their roles in hidden and output layers for classification and probability outputs.

Explore backpropagation theory and its role in training neural networks, from forward propagation and activation functions to computing mean squared error derivatives and updating weights via gradient descent.

Explore a toy three-layer neural network in PyTorch, with input, hidden, and output layers, using tensors, random weights, biases, sigmoid activation, and backpropagation with a learning rate.

This lecture sets up a PyTorch neural network with the iris data, converts labels to numbers, and trains a four-input, three-output model using relu activation and cross-entropy loss.

Prepare the iris dataset by dropping the id column, normalize predictors, and train a two-hidden-layer PyTorch neural network for iris classification using SGD.

Explore a PyTorch deep neural network for the iris dataset, encoding labels, and training a three hidden layer model with relu and dropout using data loaders and numpy.

Identify credit card fraud using a deep neural network in PyTorch, with normalization, a train-test split, and a 28-340-220-200-17-2 architecture achieving 99% test accuracy and a detailed classification report.

Explore confusion matrices and accuracy measures for binary and multi-class classification, defining true positives, true negatives, false positives, and false negatives, plus precision, recall, and Cohen's kappa.

Define what an image is by examining pixels and subunits. Explain how RGB rendering uses bits per pixel and 0 to 255 pixel values to render color images.

Learn practical techniques to read images in python using PIL, image io, and matplotlib, reading jpg and tiff files from folders, inspecting image properties, and displaying results in notebooks.

Learn basic image conversions in Python by reading an image with PIL, converting RGB to grayscale and to HSV, and visualizing the original and transformed images.

Discover how artificial intelligence, machine learning, and deep learning interrelate, with neural networks mimicking human cognition and real world applications in cancer detection and wildlife monitoring.

Build and classify images with a simple artificial neural network in PyTorch, using a 3000-input, 128-hidden, 15-class architecture trained on Fruits 360 data with 100x100 RGB images.

Build a deep neural network with two hidden layers of 200 neurons to classify 15 fruit image classes, training on an 80/20 split and achieving 100% test accuracy with SGD.

Explore cnn for image classification, using convolution, pooling, and Relu to extract features, then fully connected layers with Softmax for final differentiation, implemented with mxnet.

Build and train a convolutional neural network in PyTorch to classify 15 fruit image classes, detailing RGB 100x100 input, conv and pooling layers, and softmax outputs.

Build a convolutional neural network with three by three filters, two by two max pooling, and dense layers to classify 28 by 28 by 1 inputs, achieving 99.3% test accuracy.

Implement a transfer learning convolutional neural network with ResNet-34 in PyTorch to classify fruit images into 15 classes, resizing inputs to 224 and training across 10 epochs.

Explore data types in statistics, including categorical and numerical data, with examples such as eye color, political affiliations, weights, and counts; learn how continuous, discrete, and ordinal data inform analyses.

Transfer learning uses a pre trained model as a starting point for new tasks, reducing data and computation needs while enabling image classification with features extracted from base networks.

Sign in with a Gmail account to access Google Cloud Platform and explore free credits, pricing tools, and core services like Compute Engine, BigQuery, Cloud SQL, Dataflow, Kubernetes, and AutoML.

Enable browser-based deployment and sharing of data science projects with Posit. Create and deploy RStudio or Jupyter notebooks, connect to GitHub, and publish Shiny, Streamlit, and Dash apps for collaboration.

Explore distributed computing through the idea of a distributed computing framework that partitions tasks across nodes, enables parallelism, ensures fault tolerance, and scales with data frameworks like Hadoop and Spark.