CNN From Scratch, Using Only NumPy

Problem

It’s easy to call model.fit() in PyTorch without ever confronting what a convolution, a backward pass, or a weight update actually are. Before relying on frameworks for research work, I wanted to be certain I understood the mechanics well enough to debug them when something goes wrong – so I built a CNN with nothing but NumPy: convolution and pooling layers, backpropagation through them, and gradient descent, all written by hand.

Approach

• Implemented convolution (with stride and padding), max pooling, ReLU, and a fully connected output layer as separate classes, each with its own manual forward and backward pass.
• Derived and implemented backpropagation through the convolution operation itself, which is the part most tutorials skip – it required working out the gradient with respect to both the input and the kernel as a correlation operation.
• Trained the network on [name the dataset, e.g. MNIST] to validate that the from-scratch implementation actually learns, comparing training curves against an equivalent PyTorch model as a sanity check.

Results

• Test accuracy on [dataset]: [add %]
• Training time per epoch (pure NumPy vs. PyTorch reference): [add comparison]

What I’d improve next

• Vectorize the convolution backward pass further – the current version is correct but not fast, which is an interesting constraint in itself.
• Add batch normalization to see the from-scratch effect on training stability.