LeNet-CNN-Implementation

A NumPy-based implementation of the LeNet convolutional neural network 1998 research paper, trained on the EMNIST dataset for handwritten digit recognition.

Demo

Interact with the trained model at: https://django-app-789455248731.asia-south1.run.app/ (Deployed on Google Cloud Run)

LeNet-CNN.mp4

---

Model Architecture

This project implements the classic LeNet-5 architecture using pure NumPy, structured as follows:

Layer Type	Details	Output Shape
Input	1 × 28 × 28 grayscale image	(1, 28, 28)
Conv1	6 filters, 5×5 kernel, stride=1	(6, 24, 24)
ReLU	Activation	(6, 24, 24)
MaxPool1	2×2 pooling	(6, 12, 12)
Conv2	16 filters, 5×5 kernel, stride=1	(16, 8, 8)
ReLU	Activation	(16, 8, 8)
MaxPool2	2×2 pooling	(16, 4, 4)
Flatten	-	(256,)
FC1	Fully Connected → 120 units	(120,)
ReLU	Activation	(120,)
FC2	Fully Connected → 84 units	(84,)
ReLU	Activation	(84,)
FC3	Fully Connected → 10 units (classes)	(10,)
Softmax	(applied in loss)	(10,)

Uses ReLU activations throughout instead of tanh, and MaxPooling instead of average pooling for better gradient flow.

---

🔧 Loss Function

This project uses a combined Softmax + Categorical Cross‑Entropy Loss, implemented as a class CrossEntropyLoss, to train the LeNet model.

Components

1. Softmax Activation
2. *Negative Log-Likelihood Loss (NLLLoss)
3. Gradient for Backpropagation

CrossEntropyLoss.get Workflow

• Applies Softmax._forward() to transform logits into probabilities.

• Uses NLLLoss() to compute the scalar loss.

• Constructs dout, the gradient of loss with respect to logits, as:

  dout = prob.copy()
  dout[np.arange(N), true_labels] -= 1
  dout /= N

• Returns both loss and dout, ready for backpropagation through the network.