PyTorch Ultimate: From Basics to Cutting-Edge

Explore artificial intelligence, differentiate it from machine learning and deep learning, and review real-world examples like Google Maps, Siri, Alexa, and neural networks.

Explore how machine learning differs from classical programming, as data and target variables drive supervised, unsupervised, and reinforcement learning to build models for regression and classification.

Identify when deep learning outperforms classical machine learning as dataset size grows. See how larger neural networks gain edge with more data, leveraging GPUs and open source tools like TensorFlow.

Analyze loss functions for regression and classification, including mean squared error, mean absolute error, mean bias error, binary cross entropy, hinge loss, and multi-class cross entropy with activation choices.

Explores underfitting and overfitting in regression and classification, explains bias and variance, and shows how model complexity and noise affect training versus test performance.

Develop a neural network from scratch through theory, the dot product, and Python coding lectures implementing this directly. Follow along as you learn to understand neural networks more clearly.

Update the bias and weights by subtracting the respective loss derivatives, scaling with the learning rate, to complete a step of the optimizer.

Learn to implement a training loop from scratch: perform forward passes, compute squared losses, backpropagate gradients, update weights with an optimizer, and evaluate on test data.

Set up a neural network from scratch, configure learning rate 0.1 and 1000 iterations, create an RN model, and run the train function to verify training completes successfully.

Explore how PyTorch training loops manage model, optimizer, and loss function, detailing forward pass, backward propagation, and optimizer step to update parameters.

Build linear regression from scratch in PyTorch, computing weights and bias, generating predictions, and plotting a regression line to compare with statistical regression.

Experiment with learning rate and the number of epochs to train the model. Observe results and identify stable, effective parameter combinations.

Save a PyTorch model by writing its state_dict to a file, then load it into a new model instance using load_state_dict, verifying parameters afterward.

Explore classification concepts like confusion matrices and ROC curves, understand accuracy metrics, and learn binary, multiclass, and multi-label classifications with coding demos, templates, and solutions.

Explore how a confusion matrix describes binary classifier performance, detailing true positives and negatives, false positives and false negatives, and deriving metrics like accuracy, specificity, sensitivity, and precision.

Explore multiclass classification with the iris dataset, featuring three classes and four features, a 20% test split, and data preprocessing to float32.

Build a custom dataset class for our accustomed dataset in a multi-class exercise; you can implement it on your own or follow along with me.

Construct a simple multiclass classification network in PyTorch using two linear layers and a softmax activation, and practice implementing it within the multiclass net exercise.

Learn how to configure a multi-class neural network in PyTorch, dynamically set classes from iris data, couple cross-entropy with softmax, and train with stochastic gradient descent.

Train the model over 100 epochs with a train loader, zero gradients, perform the forward pass, compute loss with the criterion, backpropagate, and apply the optimizer step while tracking losses.

Predict the most common class as a naive baseline, yielding 40% accuracy, then compare to a simple model that reaches 93% accuracy on the test set.

Explore convolutional neural networks for image classification, covering theory and practical image preprocessing, then practice with a dataset, build an image classifier, and get layer construction tips.

Explore convolutional neural networks for computer vision, learning local patterns with filters, building feature maps through convolution and pooling, and connecting to dense layers for final predictions.

Explore how convolutional filters transform grayscale images by applying nine-cell kernels, compute results through summing products, and compare sharpen, blur, and other methods.

Explore how images are stored as three-dimensional tensors in PyTorch and how to preprocess them with resizing, center crop, grayscale, rotation, vertical flip, and normalization.

Explore binary image classification with two classes, using a concrete crack images dataset to train a convolutional neural network that detects crack or no crack with high confidence.

Engage in a multiclass image classification exercise by building a model, preparing train and test data, training with a chosen loss and optimizer, and evaluating accuracy and misclassifications.

Learn to debug a PyTorch CNN by performing a forward pass with a random 32-by-32, 3-channel input and inspecting outputs through conv, pooling, flatten, and dense layers.

Convert audio to images using spectrograms for efficient classification with image models. Apply fft to transform time-domain signals into frequency-domain representations and generate a spectrum for classification.

Prepare audio data for classification by loading wav files, extracting class types from file names, and splitting them into train and test sets while generating spectrogram images.

Demonstrates an end-to-end audio classification workflow by converting audio to grayscale 100x100 images, applying transforms, building and training a CNN, and evaluating with accuracy and a confusion matrix.