Deep Learning Masterclass with TensorFlow 2 Over 20 Projects

Explore TensorFlow indexing for 1d, 2d, and 3d tensors, using zero-based and negative indices, slicing with start:end and end+1, and selecting specific blocks by row, column, and range.

Explore TensorFlow linear algebra operations, including matmul and A@B, transpose, 3D tensor batch matrix multiplication, band_part, determinant, inverse, and einsum for efficient tensor calculations.

Explore common TensorFlow methods, including expand_dims, squeeze, reshape, concatenate, stack, gather, and gather_nd, with practical examples of axis handling, shapes, broadcasting, and padding.

Explore ragged tensors in TensorFlow, learn why non-rectangular data needs them, and create and manipulate ragged tensors with ragged constants, boolean masks, and from_row_splits.

Learn how sparse tensors efficiently represent tensors with many zeros using TensorFlow's SparseTensor, including defining indices, values, and dense_shape, and converting to dense with tf.sparse.to_dense.

Prepare data for deep learning by exploring the Secondhand Cars Dataset, examining features from years to current price, and creating shuffled inputs X and outputs Y for training.

Build a simple linear regression model in TensorFlow using a normalizer and dense layer to learn M and C, and explore Sequential, functional, and subclassing APIs with model summaries.

Explore error sanctioning in linear regression by measuring prediction errors, updating model parameters to reduce loss, and comparing mean squared error, mean absolute error, and uber loss in TensorFlow.

Explore how stochastic gradient descent optimizes model parameters by iteratively updating weights with the learning rate to reduce loss, illustrating convergence across epochs.

Split data into train, validation, and test sets; validate during training; prevent leakage by normalizing on the training data; evaluate with validation and test metrics and predict vs actual values.

Explore TensorFlow datasets with the TensorFlow data API to build efficient data pipelines, including from tensor slices, shuffling, batching, and prefetching. Apply these concepts to train, validation, and test datasets.

Prepare malaria cell images in TensorFlow 2 by loading the malaria dataset with tfds, inspecting image attributes, and splitting into train, validation, and test sets with shuffling.

Visualize images and their labels using matplotlib by creating a 4x4 grid of samples, displaying each image with its label through a get_label_string mapping for parasitized or uninfected.

Explore how convolutional neural networks reduce parameters through kernels and receptive fields, cover padding and stride effects, and compare LeNet, AlexNet, and CNN architectures with TensorFlow/Keras.

Explore binary cross entropy loss, its formula with y and y_hat, how it penalizes wrong predictions, and its use in TensorFlow with from_logits, sigmoid, and label smoothing.

Train convnets with fit on batched train and validation datasets, using batch size 32, shuffled, prefetching, and autotune; improve performance with relu, sigmoid output, batch normalization, and binary accuracy.

Evaluate the trained model on the test dataset, achieving 93% accuracy, then predict parasitized or not with a simple threshold, and visualize nine predictions to compare expected and actual results.

Learn how to save and load TensorFlow models in Google Drive, including full model vs weights, using HDF5, and continue training from saved states in Colab.

Explore the functional API in TensorFlow to build models beyond the sequential API, including multi-input multi-output architectures, ResNet-style structures, shared layers, and custom layers.

Explore model subclassing by building a feature extractor as a custom tf.keras layer, integrating conv, batch norm, and pooling into a Lenet model, and resolving parameter order for training.

Create a custom NeuralearnDense layer in TensorFlow, defining weights and biases, adding activation options (relu, sigmoid), and integrate it into a sequential model to compare initializations and training results.

Explore evaluating binary classifiers beyond accuracy by using the confusion matrix, precision, recall, roc curves, and auc to select optimal thresholds and balance false positives and false negatives.

Learn to compute and visualize a confusion matrix and ROC plot with sklearn and seaborn, compare true positives and negatives with false positives and false negatives, and adjust thresholds.

Master learning rate scheduling with TensorFlow 2 to balance speed and stability, using high initial rates, gradual decay, warm-up, and methods like exponential, cosine, and cyclical schedules, plus model checkpoints.

Learn how to implement model checkpointing in TensorFlow, save best weights based on validation loss, configure save options, and apply plateau-based learning rate reductions to improve training.

Mitigate overfitting and underfitting by applying data augmentation, dropout, regularization, early stopping, smaller networks, normalization, and hyperparameter tuning, with practical TensorFlow demonstrations.

Albumentations provides fast, unified data augmentation with a declarative pipeline in TensorFlow 2 and PyTorch, automatically adjusting bounding boxes and masks across object detection and segmentation.

Switch between eager and graph modes in TensorFlow 2 with tf.function, and learn how tracing creates portable graphs for GPU-accelerated computation.

Learn to build a custom TensorFlow 2 training loop from scratch, replacing fit with neural learn, using gradient tape for automatic differentiation, and including train and validation blocks.

Explore visualizing TensorBoard graphs for TensorFlow, compare op graph and keras conceptual graphs, and inspect conv, dense, batch normalization, and the Adam optimizer to understand computations.

Explore hyperparameter tuning with TensorBoard, using HParams to optimize learning rate, dropout, regularization, and number of units through grid search and parallel coordinates, improving model accuracy.