dependency free python implementation of mnist

mnist.py

def sigmoid(x): return 1/(1 + __import__('math').exp(-x))

def train_mnist():
    # Generate mock MNIST data (28x28 images)
    X = [[float(i%2) for i in range(784)] for _ in range(100)]  # Mock input
    y = [[1 if i==j else 0 for i in range(10)] for j in range(100)]  # Mock labels
    
    # Initialize weights and biases
    W1 = [[0.01*((i+j)%2) for j in range(784)] for i in range(30)]
    W2 = [[0.01*((i+j)%2) for j in range(30)] for i in range(10)]
    b1 = [0.0]*30
    b2 = [0.0]*10
    
    # Training loop
    for epoch in range(5):
        for i in range(len(X)):
            # Forward pass
            h = [sigmoid(sum(W1[j][k]*X[i][k] for k in range(784)) + b1[j]) for j in range(30)]
            o = [sigmoid(sum(W2[j][k]*h[k] for k in range(30)) + b2[j]) for j in range(10)]
            
            # Backward pass (simplified)
            for j in range(len(W2)):
                for k in range(len(W2[0])):
                    W2[j][k] += 0.1 * (y[i][j] - o[j]) * h[k]
            
            for j in range(len(W1)):
                for k in range(len(W1[0])):
                    W1[j][k] += 0.1 * sum((y[i][m] - o[m]) * W2[m][j] for m in range(10)) * X[i][k]

    return W1, W2, b1, b2