TensorFlow study projects

I'll use this repo to store my TensorFlow (TF) study projects. I'm following here instructions from Daniel Bourke's "TensorFlow Developer Certificate in 2022: Zero to Mastery" course on Udemy.

Project 1 - Tensor Fundamentals

1. Reviewing concepts of tensor: creation of scalar, vector and matrix tensors with "tf.constant" and "tf.Variable";
2. Getting tensor attributes (shape, rank, index, size) and performing math (e.g., "tf.math.add") and matrix operations;
3. Extra tensor operations: getting absolute, min, max and mean values. Additionally, I learn how to identify index of max value, remove size-1 dimensions from a tensor or use one-hot encoder. Extra math operations include: square, square root and logarithm;

Note: tf.config.list_physical_devices in the 3rd notebook is just to show what type of processor (CPU/GPU/TPU) we have access to.

Project 2 - Image Generator

1. Using StableDiffusion model within "KerasCV" to generate images from the natural language descriptions (text-to-image).

Project 3 - Neural Network (Regression)

1. Starting with a basic introduction to the neural networks with TF: from creating and training new model to improving its accuracy;
2. Modeling evaluation: importance of train / test data split and model structure visualisation. Calculating mean absolute error (MAE) and mean squared error (MSE);
3. Saving and loading trained NN model in 2 supported formats: "SavedModel" and "HDF5";
4. Example of NN model trained on external big dataset. Also, introduces the use of one-hot-encoder with pandas' "get_dummies", as well as normalisation ("MinMaxScaler") and standardisation ("StandardScaler").

Project 4 - Neural Network (Classification)

1. Generating dataset with SKLearn's "make_circles" package. Creating first Classification NN with a Binary Enthropy to identify which circle a feature belongs to. First 4 attempts failed, as the model expects linearity;
2. Adjusting the model, to introduce non-linearity, starting to use relu and sigmoid activations. Introducing learning rate's callback and confusion matrix;
3. Exploring multiclass classification with neural networks and visualising with Keras' "plot_model" utility:
      ==> if labels are hot endoded, expected loss function: "CategoricalCrossentropy";
      ==> if labels are integer, then expected loss function: "SparseCategoricalCrossentropy".

Project 5 - Computer Vision

1. Comparing Convolutional NN to Dense NN: less number of parameters, and still better learning about image features. Introducing "Conv2D" and "MaxPooling2D" layers;
2. Creating baseline CNN model for the "Binary Classification", reducing overfitting through data augmentation. Preparing custom image for the use by CNN model, introducing "tf.io.read_file" and "tf.image";
3. Creating baseline CNN model for the "Multiclass Classification", introducing "tf.keras.models.clone_model".

Project 6 - Transfer Learning

1. Using "tensorflow_hub" package to pull trained models from the TensorFlow Hub repository and utilise TF Sequential API to embed them into our custom models as a part of Transfer Learning's "Feature Extraction". Introducing callbacks and comparing performance of different models via "TensorBoard";
2. Utilising TF Functional API to combine pre-trained model and fine-tune it for specific scenario. Saving training-learnt weights with Checkpoint callback. Fine-tuning the model, by making some of the base model's layers retrainable;
3. Scaling the model up from 10 to 101 classes. Using SKLearn's classification reports to evaluate model efficiency.

Project 7 - Food Vision (Milestone 1)

1. Introducing TensorFlow Datasets (TFDS). Learning about batching and prefetching. Creating modelling callbacks.

Project 8 - Natural Language Processing (NLP)

1. Using "Recurrent Neural Networks (RNN)" to build NLP models. Converting text to numbers with "Tokenisation" (Text Vectorisation) and "Embedding". Visualising embedding layer's learnt weights with the TensorFlow Embedding Projector. Building RNN with LSTM- (Long Short-Term Memory) and GRU- (Gated Recurrent Unit) cells. Measuring prediction run time.

Project 9 - Text Abstracts (Milestone 2)

1. Using RNN to skim medical texts and generate output text as per expected structured format. Introducing multimodal setup with layers.Concatenate.

Project 10 - Time Series

1. Using time series analysis to predict Bitcoin prices. Applying Mean Absolute Scaled Error (MASE) metrics. Introducing terms of window and horizon. Building various models to beat autocorrelation base model. Re-creating N-BEATS algorithm.

Project 11 - Covid Results Predictions

1. Results of the Cambridge Spark's Marc 2023 Hackathon. Predicting how likely the patient to get positive or negative Covid results, based on their answers from the provided questionnaire. Worked on the adapted dataset from the Israeli Ministry of Health - source is referenced within the Jupyter notebook.