This is the code I wrote for a Convolutional Neural Network from scratch, purely in NumPy and using only arrays. The only exceptions are when I used Keras to load the MNIST dataset and to turn my output, into a categorical numpy array. Other than this, the math and everything else was self-coded, with inspiration for some functions, taken from various GitHub repos.

This model is built like how Tensorflow/Keras's Sequential models work, taking in a list of layer classes and computing the forward propagation and backpropgation algorithms within the layers, and storing the parameters within the layers themselves.

---

Architecture of the CNN made

Training data: (160, 28, 28)

Per Image: (28, 28)

Per image network architecture:

• Convolution - (28, 28) to (14, 14, 2)
  • Filter size: (2x2)
  • Stride: 2
  • No of kernels = no of digits to be recognized (in this case, 2)
• Activation - Sigmoid
• MaxPooling - (14, 14, 2) to (7, 7, 2)
  • Stride: 1
  • Filter size: (2x2)
  • Pad: None
• Reshape - (7, 7, 2) to (98, 1) (flatten operation)
• Dense - 98 to 35
• Activation - Softmax
• Dense - 35 to 2
• Activation - Softmax
• Loss - Cross Entropy