Want genuine suggestion to build Support Vector Ma

Question

As i know how to build a Support Vector Machine using Scikit-Learn but now i want to make it from scratch in python without using Scikit-Learn. As i am confused and having lack of knowledge about the internal processes i would be greatly pleased if getting help and make it out.

Answer 1

You can implement a simple linear SVM with numpy only like below. BTW, please google before you ask question next time. There are lots of resources and tutorial online.

    import numpy as np
    def my_svm(dataset, label):
        rate = 1 # rate for gradient descent
        epochs = 10000 # no of iterations
        weights = np.zeros(dataset.shape[1]) # Create an array for storing the weights

        # Min. the objective function(Hinge loss) by gradient descent
        for epoch in range(1,epochs):
            for n, data in enumerate(dataset):
                if (label[n] * np.dot(dataset[n], weights)) < 1:
                    weights = weights + rate * ( (dataset[n] * label[n]) + (-2  *(1/epoch)* weights) )
                else:
                    weights = weights + rate * (-2  * (1/epoch) * weights)

        return weights

    def predict(test_data,weights):
        results = []
        for data in test_data:
            result = np.dot(data,weights)
            results.append(-1 if result < 0 else 1)
        return results

Generate dataset for training and testing

    dataset = np.array([
        [-2, 4,-1], #x_cood,y_cood,bias
        [4, 1, -1],
        [0, 2, -1],
        [1, 6, -1],
        [2, 5, -1],
        [6, 2, -1]
        ])
    label = np.array([-1,-1,-1,1,1,1])

    weights = my_svm(dataset,label)

Test it

    test_data = np.array([
                [0,3,-1], #Should belong to -1
                [4,5,-1]  #Should belong to 1
                ])
    predict(test_data, weights)
    >Out[10]: [-1, 1]