Deep Learning : Image Classification with Tensorflow in 2025

Explore tensor basics—from ranked tensors, to zero- and one-dimensional, to three-dimensional tensors—covering shapes, indexing, broadcasting, casting, and matrix operations in deep learning.

Learn to create and cast tensors in TensorFlow from 0D to 4D, convert numpy arrays, and generate random normal and uniform tensors with seed for reproducibility.

Explore TensorFlow indexing across 1D, 2D, and 3D tensors, using zero-based and negative indices, slices with colon, steps, and ellipses to retrieve values and subarrays.

Explore TensorFlow math operations for tensors, covering absolute, add, divide; divide no nan; broadcasting; max; min; arg max; equal; power; reduce methods; top k; and sigmoid.

Explore linear algebra operations in TensorFlow linalg, including matrix multiplication, transpose, insum, and batched 3D tensor multiply, with practical shape and broadcasting insights.

Explore TensorFlow tensor manipulation using common methods such as expanddims, squeeze, reshape, concatenate, stack, pad, and gather and gather_nd, with axis handling and batch shape transformations.

Explore ragged tensors in TensorFlow, learning to convert non rectangular Python sequences into ragged tensors using row splits and boolean masks to manage non uniform data.

Explore how TensorFlow represents sparse data with sparse tensors using indices, values, and a dense shape to store many zeros efficiently, and enable sparse to dense conversion for computation.

Explore how to handle strings in TensorFlow using the strings module, creating 1D string tensors, and performing operations like join, length, and string conversion for model workflows.

Learn how variables in TensorFlow maintain a shared state defined by an initial tensor, with the type and shape fixed, and update values using assign methods like assign_add and assign_sub.

Learn to build a linear regression model to predict car prices using multiple features, following the machine learning life cycle from task definition to training and evaluation.

Prepare car price data for model training by shuffling, splitting into inputs and outputs, reshaping tensors, and applying adaptive normalization with TensorFlow, using Kaggle's second hand cars dataset.

learn to build simple machine learning models with TensorFlow using linear regression y = m x + c, normalization and dense layers, and explore model summaries and trainable parameters.

Learn to measure and reduce regression error using mean squared error, mean absolute error, and Uber loss in TensorFlow and Keras, by comparing y true and y pred.

Explore training and optimization of a neural model using stochastic gradient descent, learning rate tuning, and loss minimization, with practical steps in TensorFlow, SGD variants, and model convergence.

Define and use the root mean square error to measure and compare model performance in TensorFlow, including RMSE in compilation, monitoring during training, and evaluation with X and y.

Split data into train, validation, and test sets to prevent leakage, evaluate unseen data, and apply normalization only on the train set during TensorFlow training.

Explore corrective measures for underfitting by analyzing loss plots, increasing model depth, and adding non-linear activations like ReLU, sigmoid, and tanh to improve TensorFlow image classification performance.

Explore TensorFlow datasets and the data API to build efficient pipelines, using from_tensor_slices, shuffle, batch, and prefetching for train, val, and test sets.

learn to build a convolutional neural network to classify blood cell images for malaria as parasitic or uninfected, preparing data, training, evaluating, and applying RGB pixel normalization.

Explore the malaria image dataset in tfds, inspect features and labels, then split data into train, validation, and test sets (80/10/10) with shuffling and loading via TensorFlow data sets.

Visualize the first 16 train samples in a 4x4 matplotlib grid, showing each image with its label as the title and mapping 0 to parasitized and 1 to uninfected.

Resize all inputs to 224 by 224 and normalize to the 0–1 range before feeding data into the model, using TensorFlow image resize and optional mean-std standardization.

Explore how convolutional neural networks enable image classification by learning kernels for 2D and 3D inputs, covering receptive fields, padding patterns, stride, pooling, and TensorFlow Keras conv2d.

Build a convolutional neural network with a sequential API by defining a 5x5x6 input, applying conv, sigmoid activation, max pooling, flatten, and dense layers for a binary two-output classifier.

Explore binary cross entropy loss for binary classification, linking y true and model predictions, illustrate with TensorFlow Keras, from_logits, sigmoid activation, and label smoothing.

Train the model with batched train and validation data, batch size 32, using shuffle, prefetch, and auto-tune, then apply relu, a single sigmoid output, and batch normalization to improve accuracy.

Evaluate the trained model on the test data to achieve about 93% accuracy, and predict parasite or not using model.predict. Visualize nine predictions to compare expected versus actual results.

Learn how to save and load TensorFlow models in Google Drive or Colab, including full model and weights, with the HDF5 format and checkpoints to resume training.

Master the functional API in TensorFlow to build multi-input, multi-output models with shared and custom layers, and learn to mix functional, sequential, and subclassed approaches.

Master model subclassing by building a TensorFlow feature extractor as a custom layer with init and call methods, integrating filters, kernel size, strides, activation, pool size, and the functional API.

Learn how to create a custom dense layer in TensorFlow by defining weights and biases, handling input shapes, and integrating it into a sequential model with activation.

Explore how to compute true positives, false positives, true negatives, and false negatives, and use precision, recall, and AUC with ROC curves to set thresholds.

Learn to evaluate a binary image classifier with confusion matrices, thresholds, and seaborn visualizations, comparing true positives, true negatives, false positives, and false negatives.

Explore how to plot the roc curve by computing false positives, true positives, and thresholds, then select a threshold to minimize false positives while preserving the true positive rate.

Master TensorFlow callbacks—from custom callbacks to csv logger and early stopping—to monitor training history and optimize learning rate and weights for image classification.

Learn how to implement learning rate scheduling in TensorFlow to balance speed and stability, using callbacks, warm up, cosine and stepwise decay, and model checkpointing.

Configure checkpointing in TensorFlow to save best weights by validation loss, manage file paths, and enable plateau-based learning rate reduction with callbacks.

Explore strategies to combat overfitting and underfitting, including data augmentation, dropout, regularization, early stopping, smaller networks, hyperparameter tuning, and normalization, with practical TensorFlow examples.

Explore data augmentation in TensorFlow using tf.image and Keras layers, applying rotations, flips, brightness, saturation, and central crop, and embedding augmentations into training pipelines for efficient model learning.

Learn mixup data augmentation in TensorFlow by blending two images with a beta-distributed lambda, updating labels, and integrating the technique into a tf.data pipeline with preprocessing.

Implement cutmix data augmentation with TensorFlow 2 and tf.data to improve image classification. Learn how to crop random patches from one image, paste onto another, and adjust labels accordingly.

Learn to implement data augmentation with albumentations in TensorFlow and PyTorch, building a reusable transform pipeline that handles bounding boxes for object detection and masks for segmentation.

Learn to build custom losses and metrics in TensorFlow 2, from simple functions to class-based implementations, including a binary cross entropy loss and custom accuracy metrics.

Switch between eager and graph modes in TensorFlow 2, understand tracing, and leverage the tf function decorator to enable portable graph execution.

Build a custom training loop in TensorFlow 2, replacing the feed method with a gradient tape, loss computation, and optimizer updates for training and validation.

Learn to log training data with the tensorboard callback, visualize model graphs on a web interface, and tune hyperparameters using logs, distributions, histograms, ROC plots, confusion matrices, and profiling.

Understand how TensorFlow builds and visualizes graphs, from operation graphs to a keras conceptual graph, using tensorboard to inspect dense layers, batch normalization, dropout, and the Adam optimizer.

Explore hyperparameter tuning in TensorFlow using TensorBoard hparams: define, sweep, and evaluate choices like dropout, regularization, units, and learning rate with grid and random search.

Explore profiling and visualizations in tensorboard to identify input time as the bottleneck, then use input pipeline analyzer, trace viewer, and tf data bottlenecks to examine prefetch and 16-bit operations.