TensorFlow-Keras using Mnist Dataset

TensorFlow is one of the popular engines for deep learning.

Why is it called TensorFlow?

The structure we use to hold our data, i.e. the scalars, the list the array, and the matrices, are all mathematically called Tensors. Deep Neural or Artificial neural networks (ANN) are all about algebra. Here the data flow in the form of a matrix from the starting initial layer to the final layer. hence the name - TensorFlow.

The beauty of this is we can do A-Z of machine learning on this. We need not employ different tools to do different parts of our modeling. Here everything I available under one single roof. It comes with the facility of distributed training. It has a very easy syntax and training becomes faster. It comes with support for GPU, provided we make use of NVIDIA library the CUDA's. Not all vendor chips are supported, it runs in most of the operating system, we are aware of. It supports Python and also other languages. many of the libraries of Tensoflow are written in C++.

Note: If you are using TensorFlow 2.0, always use Tensorflow Keras.

It also contains the TF function, TF Gradient Tape, and TF data. If I want to generate the graph in TF 2.0, then we will use the TF function.

Gradient Tape: When we do the Back Propagation, the chain equation, TensorFlow 1.0 used to make use of the graph and then goes reverse in the graph doing dy/dx, assuming all the nodes in the graph are derivable. Tensorflow 2.0 does not have any graph, so it cannot do an automatic dy/dx in the chain equation.

So how will we do this backpropagation there?

It gives a facility called Gradient Tape. The Gradient Tape just records all the steps. It keeps a tape recorder of all the steps and uses it to o the backpropagation automatically. What used to be done by a graph in Tensrflow 1.0 is done by a recorder in 2.0

Just like a pipeline in scikit-learn we can create a pipeline of dataflow using the TF data function in Tensorflow. We use model building sequential where the layers are added in sequence; we can also use a functional approach, which is similar but more versatile. TensorFlow gives us facilities to customize your own layers, your own metrics, and your own loss functions.

You can modify every class of TensorFlow that is available. they have given the entire source code.

Keras: It is a high-level API that we make use of to write our programs in TensorFlow. Keras is a high-level neural network API, written and implemented in Python. It can run on top of TensorFlow. It was designed keeping fast experimentation in mind.

Tensorflow Keras _ DeepLearningMnist.ipynb <https://colab.research.google.com/drive/1vNjPzZpx4CwKCb_ebVeEm8SDzUjDewTb?usp=sharing>