Keras Basics — Models, Layers, and Tensors

Keras models are built from layers, and layers operate on tensors.

What a Tensor Is

A tensor is just an N-dimensional array (a generalisation of scalars, vectors, and matrices).

Examples:

• Scalar → 0D tensor
• Vector → 1D tensor
• Matrix → 2D tensor
• Image → 3D tensor (height × width × channels)
• Batch of images → 4D tensor (batch × H × W × C)

In Keras/TensorFlow, your input data will always be a tensor or converted into one.

You don’t usually work with tensors manually, Keras handles it for you, but you must know your input shape.

What a Layer Is

A layer is a transformation applied to data.

Examples:

• Dense layer → uses weights + bias to turn one tensor into another
• Conv2D layer → performs convolution on images
• LSTM layer → processes sequences
• Dropout → randomly drops activations to reduce overfitting
• BatchNormalization → stabilises training
• Flatten → converts 2D/3D tensors into 1D tensors

Each layer takes a tensor, applies a transformation, outputs another tensor.

What a Model Is

A Keras model is just a connected stack (or graph) of layers.

Two main types:

• Sequential API → simple stacks
• Functional API → flexible graphs

Sequential API (Simple Models)

Use Sequential when the architecture is linear:

Example: Simple feed-forward network

Output:

Output explanation:

• Each layer's type
• Output shape
• Number of parameters
• Total trainable parameters

Sequential is perfect for:

• Small MLPs
• Basic CNNs
• Many educational models

But it cannot do:

• Multi-input networks
• Multi-output networks
• Skip connections
• Complex architectures

Functional API (Flexible and Modern)

Functional API is recommended for most real projects.

Example:

Output:

Why use Functional API?

• You can create flexible, non-linear topologies
• You can reuse layers
• You can build multi-branch / multi-output models
• You can build architectures like ResNet, U-Net, Transformers

Input Shapes

You must know the shape of your data.

Some examples from test data:

• MNIST image (28×28 grayscale) - (batch, 28, 28)
• CIFAR-10 image (32×32 RGB) - (batch, 32, 32, 3)
• Flattened MNIST - (batch, 784)

Almost all training bugs come from incorrect input shapes.

Layers: The Most Common Types

Dense (Fully Connected)

Used for:

• Classification (MLP)
• Regression
• Final layers in CNNs / RNNs

Convolutional Layers (Images)

Recurrent Layers (Sequences)

Preprocessing Layers

Regularisation

Model Summary

Every time you build a model:

You get:

• Layer names
• Output shapes
• Parameter counts
• Total number of weights

This instantly tells you:

• Did you pass the right input shape?
• Did a Flatten layer go missing?
• Does the architecture look correct?

Example: End-to-End Mini Model with Real Data

We won't train it fully yet, this is just to show shape and flow.

Using MNIST data:

Build model:

Check structure:

Output:

This confirms the tensor shapes flow correctly.

Debugging Common Shape Errors

Error: Expected shape (None, 784) but got (None, 28, 28)

Fix:

Add a flatten layer:

Error: Input shape mismatch between Functional layers

Fix:
Ensure you apply layers to the same tensor objects.