[Python] Hey guys,trying to buid a Convolutional Neural Network using python from scratch

import numpy as np
import matplotlib.pyplot as plt

def OneHot(y):
return np.eye(np.max(y.astype(int)) + 1)[y.astype(int)]

def CE(Y, Y_hat):
epsilon = 1e-10
Y_hat = np.clip(Y_hat, epsilon, 1 - epsilon)
return -np.mean(np.sum(Y * np.log(Y_hat), axis=0))

def CE_prime(Y, Y_hat):
return Y_hat - Y

class Dense:
def __init__(self, n_inputs, n_outputs):
self.weights = np.random.randn(n_outputs, n_inputs) * 0.01
self.biases = np.zeros((n_outputs, 1))

def forward(self, X):
self.X = X
self.Z = np.dot(self.weights, X) + self.biases
return self.Z

def backward(self, grad, lr):
dZ = grad
dW = np.dot(dZ, self.X.T) / self.X.shape[1]
db = np.sum(dZ, axis=1, keepdims=True) / self.X.shape[1]
dX = np.dot(self.weights.T, dZ)

self.weights -= lr * dW
self.biases -= lr * db

return dX

class Conv2D:
def __init__(self, kernel_size, input_channels, input_shape):
self.kernel_size = kernel_size
self.input_channels = input_channels
self.input_shape = input_shape
self.weights = np.random.randn(input_channels, kernel_size, kernel_size) * 0.01
self.biases = np.zeros(1)

def forward(self, X):
self.X = X
batch_size,_, H, W = X.shape
output_H = H - self.kernel_size + 1
output_W = W - self.kernel_size + 1
self.Z = np.zeros((batch_size,output_H, output_W))

for b in range(batch_size):
for i in range(output_H):
for j in range(output_W):
patch = X[b, :, i:i+self.kernel_size, j:j+self.kernel_size]
self.Z[b, i, j] = np.sum(self.weights * patch) + self.biases
return self.Z

def backward(self, grad, lr):
batch_size, output_H, output_W = grad.shape
_, _, H, W = self.X.shape
kernel_size = self.kernel_size

dW = np.zeros_like(self.weights)
db = np.sum(grad) / batch_size
dX = np.zeros_like(self.X)

for b in range(batch_size):
for i in range(output_H):
for j in range(output_W):
patch = self.X[b, :, i:i+kernel_size, j:j+kernel_size]
dW += grad[b, i, j] * patch / batch_size
dX[b, :, i:i+kernel_size, j:j+kernel_size] += self.weights * grad[b, i, j]

self.weights -= lr * dW
self.biases -= lr * db

return dX

class MaxPool2D:
def __init__(self, pool_size):
self.pool_size = pool_size

def forward(self, X):
self.X = X
batch_size, H, W = X.shape
output_H = H // self.pool_size
output_W = W // self.pool_size
self.Z = np.zeros((batch_size, output_H, output_W))
self.max_indices = []

for b in range(batch_size):
for i in range(output_H):
for j in range(output_W):
h_start = i * self.pool_size
h_end = h_start + self.pool_size
w_start = j * self.pool_size
w_end = w_start + self.pool_size
patch = X[b, h_start:h_end, w_start:w_end]
max_val = np.max(patch)
self.Z[b, i, j] = max_val
max_idx = np.unravel_index(np.argmax(patch), patch.shape)
self.max_indices.append((b, h_start + max_idx[0], w_start + max_idx[1]))

return self.Z

def backward(self, grad,lr):
dX = np.zeros_like(self.X)
grad_flat = grad.flatten()
for idx, (b, h, w) in enumerate(self.max_indices):
dX[b, h, w] += grad_flat[idx]
return dX

class Flatten:
def forward(self, X):
self.input_shape = X.shape
batch_size = X.shape[0]
return X.reshape(batch_size, -1).T

def backward(self, grad, lr):
return grad.T.reshape(self.input_shape)

class Relu:
def forward(self, X):
self.X = X
return np.maximum(X, 0)

def backward(self, grad, lr):
return grad * (self.X > 0)

class Softmax:
def forward(self, X):
exp_X = np.exp(X - np.max(X, axis=0, keepdims=True))
self.output = exp_X / np.sum(exp_X, axis=0, keepdims=True)
return self.output

def backward(self, grad, lr):
return grad

class NeuralNetwork:
def __init__(self, layers=[]):
self.layers = layers

def add(self, layer):
self.layers.append(layer)

def train(self, epochs, X, Y, lr, batch_size=32):
num_samples = X.shape[1]
for epoch in range(epochs):
permutation = np.random.permutation(num_samples)
X_shuffled = X[:, permutation]
Y_shuffled = Y[:, permutation]

for i in range(0, num_samples, batch_size):
X_batch = X_shuffled[:, i:i+batch_size]
Y_batch = Y_shuffled[:, i:i+batch_size]

Y_hat = X_batch
for layer in self.layers:
Y_hat = layer.forward(Y_hat)

loss = CE(Y_batch, Y_hat)
grad = CE_prime(Y_batch, Y_hat)

for layer in reversed(self.layers):
grad = layer.backward(grad, lr)

if epoch % 1 == 0:
print(f"Epoch [{epoch}/{epochs}] Loss: {loss:.5f}")

def score(self, X, Y):
Y_hat = self.predict(X)
Y_true = np.argmax(Y, axis=0)
return np.mean(Y_hat == Y_true) * 100

def predict(self, X):
Y_hat = X
for layer in self.layers:
Y_hat = layer.forward(Y_hat)
return np.argmax(Y_hat, axis=0)


from sklearn.model_selection import train_test_split
from sklearn.datasets import load_digits
import numpy as np

digits = load_digits()
X = digits.images
y = digits.target

X = X

X = X.reshape(-1, 1, 8, 8)

Y = OneHot(y)

X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.2)
Y_train, Y_test = Y_train.T , Y_test.T

nn = NeuralNetwork()
nn.add(Conv2D(kernel_size=3, input_channels=1, input_shape=(1, 8, 8)))
nn.add(Relu())
nn.add(MaxPool2D(pool_size=2))
nn.add(Flatten())
nn.add(Dense(9, 24))
nn.add(Relu())
nn.add(Dense(24, 10))
nn.add(Softmax())

epochs = 5
learning_rate = 5
nn.train(epochs, X_train, Y_train, learning_rate, batch_size=1)

accuracy = round(nn.score(X_test, Y_test),2)
print(accuracy)


So this converges really slow with 0.01 and 0.1 learning rate,also i think loss works well and it converges to almost zero.But accuracy is around 10%.Normal Dense layer neural network works fine(like iris dataset).I asked AI's like chatgpt, deepseek and claude but they just make it worse.What should i do?

Continue reading...