I coded the perceptron algorithm form scratch and compared the weights I obtained after training with the weights I obtained after training the sklearn perceptron model. I believe even the sklearn model initializes the weights and biases as a zero vector and I chose the learning rate eta0=1 to match my perceptron code. (Note: the bias in my code is the last term in the vector w_b)

My Code:

def perceptron(X_train, y_train):
    #initialize weights as 0
    w = np.zeros(len(X_train.columns))
    b = 0
    w_b = np.append(w, b)
    while True:
        misclassifications = 0  
        for X , Y in zip(X_train.values, y_train.values):
            X_i = np.append(X, 1)
            if Y*(np.dot(X_i,w_b)) <= 0:
                w_b = w_b + Y*X_i
                misclassifications += 1
        if misclassifications == 0:
            break
    return w_b

w_b = perceptron(X_train, y_train)

Result: [-3. 6.7 -1. ]

sklearn code:

perceptron = Perceptron(max_iter=1000, eta0=1,random_state=42) 
perceptron.fit(X_train, y_train)

print("weights are",perceptron.coef_)
print("bias is",perceptron.intercept_)

Result: weights are [[-4.7 10.1]] bias is [-2.]

I expected the weights to be same but they aren’t. Any clue on why?