感知机算法的pytorch实现代码

Question

这是我的代码

import matplotlib.pyplot as plt
import torch
import torch.utils.data as Data
import numpy as np


class Perceptron:
    def __init__(self, X, y, learn_rate=0.01, batch_size=32, epoch=100):  # feature_num特征数，label_num标签数，dataMat训练数据矩阵
        self.feature_num = X_train.shape[1]
        self.label_num = 1
        self.weight = torch.normal(0, 0.01, size=(self.feature_num,), requires_grad=True) 
        self.bias = torch.normal(0, 0.01, size=(1,), requires_grad=True)
        self.train_iter = Data.DataLoader(
            Data.TensorDataset(torch.tensor(X, dtype=torch.float32), torch.tensor(y, dtype=torch.float32)),
            batch_size=batch_size,
            shuffle=True)
        self.batch_size = batch_size
        self.learn_rate = learn_rate
        self.epoch = epoch

    def loss(self, X, y):  # 单次loss function的相反数
        return - torch.mul(y, (torch.matmul(X, self.weight) + self.bias))

    def train(self):  # 训练函数
        log = []
        for j in range(self.epoch):
            all_l = 0
            for X, y in self.train_iter:
                l = self.loss(X, y).sum()
                # 反向传播
                l.backward()
                # 梯度更新
                self.weight.data = self.weight.data + self.learn_rate * self.weight.grad.data
                self.bias.data = self.bias.data + self.learn_rate * self.bias.grad.data
                self.weight.grad.data.zero_()
                self.bias.grad.data.zero_()
                all_l += l.data.item()
            log.append(all_l / 100000)
        plot_history(log)
        return log

    def predict(self, X, y=None):
        X = torch.tensor(X, dtype=torch.float32)
        pred_num = X.shape[0]
        ans = []
        for i in range(pred_num):
            ans.append(torch.sign(torch.dot(self.weight, X[i])))
        if y is not None:
            y = torch.tensor(y, dtype=torch.float32)
            plot(X, y, self.weight, self.bias)
        return ans


def generate():
    from sklearn.datasets import make_blobs
    X_data, y_data = make_blobs(n_samples=1000, n_features=2, centers=2)
    X_train, y_train = X_data[:800], y_data[:800]
    X_test, y_test = X_data[800:], y_data[800:]
    y_train = np.where(y_train == 0, -1, 1)
    y_test = np.where(y_test == 0, -1, 1)
    return X_train, y_train, X_test, y_test


def plot(X, y, w, b):
    with torch.no_grad():
        plt.scatter(X[:, 0], X[:, 1], c=y)
        x = torch.linspace(-10, 10, 500)  # 创建分类线上的点，以点构线。
        y = -w[0] / w[1] * x - b / w[1]
        plt.scatter(x, y, c=torch.zeros(size=(500,)))
        plt.show()


def plot_history(history):
    plt.plot(np.arange(len(history)), history)
    plt.show()


X_train, y_train, X_test, y_test = generate()
model = Perceptron(X_train, y_train)
model.train()
model.predict(X_test, y_test)

感觉有很多问题，pytorch也不熟悉，希望有人指点迷津

Answer 1

应该解决了

import matplotlib.pyplot as plt
import torch
import torch.utils.data as Data
import numpy as np


class Perceptron:
    # 注意,为了我们能看到训练的效果，特意将learning_rate设的很小
    def __init__(self, X, y, learn_rate=0.00001, batch_size=16, epoch=100):  
        # feature_num特征数，label_num标签数
        self.feature_num = X_train.shape[1]
        self.label_num = 1
        # 权重初始化为均值为1，方差为0.01的正态随机数
        self.weight = torch.normal(1, 0.01, size=(self.feature_num,), requires_grad=True)
        # 偏差初始化为均值为0，方差为0.01的正态随机数
        self.bias = torch.normal(0, 0.01, size=(1,), requires_grad=True)
        # 批数据生成器
        self.train_iter = Data.DataLoader(
            Data.TensorDataset(torch.tensor(X, dtype=torch.float32), torch.tensor(y, dtype=torch.float32)),
            batch_size=batch_size,
            shuffle=True)
        self.batch_size = batch_size
        self.learn_rate = learn_rate
        self.epoch = epoch

    def loss(self, X, y):
        l = - torch.mul(y, (torch.matmul(X, self.weight) + self.bias))
        # 损失函数只计算错误分类的样本，故l小于0时应当作做0处理，相当于使用ReLU做处理
        return torch.nn.ReLU()(l)

    def train(self):  # 训练函数
        log = []
        for j in range(self.epoch):
            all_l = 0
            for X, y in self.train_iter:
                l = self.loss(X, y).sum()
                if l > 0:
                    # self.weight = self.weight + self.learn_rate * torch.matmul(y, X)
                    # self.bias = self.bias + self.learn_rate * y.sum()
                    l.backward()
                    self.weight.data = self.weight.data - self.learn_rate * self.weight.grad.data
                    self.bias.data = self.bias.data - self.learn_rate * self.bias.grad.data
                    self.weight.grad.data.zero_()
                    self.bias.grad.data.zero_()
                    all_l += l.data.item()
            log.append(all_l)
            # plot(X_test, y_test, self.weight, self.bias)
        return log

    def predict(self, X, y=None):
        X = torch.tensor(X, dtype=torch.float32)
        pred_num = X.shape[0]
        ans = []
        for i in range(pred_num):
            ans.append(torch.sign(torch.dot(self.weight, X[i])))
        if y is not None:
            y = torch.tensor(y, dtype=torch.float32)
            print("识别正确率：%s"%((torch.tensor(ans) == y).sum()/y.shape[0]))
            plot(X, y, self.weight, self.bias)
        return ans


def generate():
    from sklearn.datasets import make_blobs
    X_data, y_data = make_blobs(n_samples=1000, n_features=2, centers=2)
    X_train, y_train = X_data[:800], y_data[:800]
    X_test, y_test = X_data[800:], y_data[800:]
    y_train = np.where(y_train == 0, -1, 1)
    y_test = np.where(y_test == 0, -1, 1)
    return X_train, y_train, X_test, y_test


def plot(X, y, w, b):
    with torch.no_grad():
        plt.scatter(X[:, 0], X[:, 1], c=y)
        x = torch.linspace(-10, 10, 500)  # 创建分类线上的点，以点构线。
        y = -w[0] / w[1] * x - b / w[1]
        plt.scatter(x, y, c=torch.zeros(size=(500,)))
        plt.show()


def plot_history(history):
    plt.plot(np.arange(len(history)), history)
    plt.show()


X_train, y_train, X_test, y_test = generate()
model = Perceptron(X_train, y_train, epoch=100)
log = model.train()
plot_history(log)
model.predict(X_test, y_test)