MnistCnn

Constructs a Convolutional Neural Network (CNN) tailored for the MNIST dataset using a DSL-based network builder.

This model consists of two convolutional blocks followed by a flattening stage and two dense (fully connected) layers. It is designed to classify handwritten digits (0–9) from grayscale 28x28 pixel images.

The architecture is as follows:

• Stage: "conv1"

• 2D Convolution with:

• 16 output channels

• 5x5 kernel

• stride of 1

• padding of 2

• ReLU activation

• 2x2 MaxPooling with stride of 2

• Stage: "conv2"

• 2D Convolution with:

• 32 output channels

• 5x5 kernel

• stride of 1

• padding of 2

• ReLU activation

• 2x2 MaxPooling with stride of 2

• Stage: "flatten"

• Flattens the tensor for dense layer input

• Stage: "dense"

• Fully connected layer with 128 units

• ReLU activation

• Stage: "output"

• Fully connected layer with 10 output units (for 10 MNIST classes)

• Softmax activation over dimension 1 to produce class probabilities