网站页面设计 颜色 背景 要求,wordpress如何关注博客,汕头网站时优化,分类信息网站做淘客看到一个有意思的题目#xff1a;仅使用python的标准库#xff0c;完成一个小批量梯度下降的线性回归算法
平常使用numpy这样的计算库习惯了#xff0c;只允许使用标准库还有点不习惯#xff0c;下面就使用这个过程来写一个。
import random
from typing import List# 生…看到一个有意思的题目仅使用python的标准库完成一个小批量梯度下降的线性回归算法
平常使用numpy这样的计算库习惯了只允许使用标准库还有点不习惯下面就使用这个过程来写一个。
import random
from typing import List# 生成测试数据
def generate_data(num_samples: int, weights: List[float], bias: float, noise0.1) - (List[List[float]], List[float]):X [[random.uniform(-10, 10) for _ in range(len(weights))] for _ in range(num_samples)]y [sum(w * x for w, x in zip(weights, x_i)) bias random.uniform(-noise, noise) for x_i in X]return X, y# 计算损失
def mse(y_true: List[float], y_pred: List[float]):return 0.5 * sum((yt - yp) for yt, yp in zip(y_true, y_pred)) ** 2# 将矩阵转置
def transpose(mat: List[List[float]]):row, col len(mat), len(mat[0])# 固定列访问行result [[mat[r][c] for r in range(row)] for c in range(col)]return result# 计算矩阵乘法
def matmul(mat: List[List[float]], vec: List[float]):return [sum(r * c for r, c in zip(row, vec)) for row in mat]# 计算梯度
def compute_grad(y_true_batch: List[float], y_pred_batch: List[float], x_batch: List[List[float]]):batch_size len(y_true_batch)residual [yt - yp for yt, yp in zip(y_true_batch, y_pred_batch)]# 根据 y x w b# grad_w -x.T residualgrad_w matmul(transpose(x_batch), residual)grad_w [-gw / batch_size for gw in grad_w]grad_b -sum(residual) / batch_size# grad_w: List[float]# grad_b: floatreturn grad_w, grad_b# 开启训练
def train():lr 0.01epochs 50batch_size 16dim_feat 3num_samples 500weights [random.random() * 0.1 for _ in range(dim_feat)]bias random.random() * 0.1print(original params)print(w:, weights)print(b:, bias)X, y generate_data(num_samples, weights, bias, noise0.1)for epoch in range(epochs):for i in range(0, num_samples, batch_size):x_batch X[i:ibatch_size]y_batch y[i:ibatch_size]y_pred [item bias for item in matmul(x_batch, weights)]loss mse(y_batch, y_pred)grad_w, grad_b compute_grad(y_batch, y_pred, x_batch)weights [w - lr * gw for w, gw in zip(weights, grad_w)]bias - lr * grad_bprint(fEpoch: {epoch 1}, Loss {loss:.3f})print(trained params)print(w:, weights)print(b:, bias)train()输出结果如下
original params
w: [0.04845598598148951, 0.007741816562531545, 0.02436678108587098]
b: 0.01644073086522535
Epoch: 1, Loss 0.000
Epoch: 2, Loss 0.000
Epoch: 3, Loss 0.000
Epoch: 4, Loss 0.000
Epoch: 5, Loss 0.000
Epoch: 6, Loss 0.000
Epoch: 7, Loss 0.000
Epoch: 8, Loss 0.000
Epoch: 9, Loss 0.000
Epoch: 10, Loss 0.000
Epoch: 11, Loss 0.000
Epoch: 12, Loss 0.000
Epoch: 13, Loss 0.000
Epoch: 14, Loss 0.000
Epoch: 15, Loss 0.000
Epoch: 16, Loss 0.000
Epoch: 17, Loss 0.000
Epoch: 18, Loss 0.000
Epoch: 19, Loss 0.000
Epoch: 20, Loss 0.000
Epoch: 21, Loss 0.000
Epoch: 22, Loss 0.000
Epoch: 23, Loss 0.000
Epoch: 24, Loss 0.000
Epoch: 25, Loss 0.000
Epoch: 26, Loss 0.000
Epoch: 27, Loss 0.000
Epoch: 28, Loss 0.000
Epoch: 29, Loss 0.000
Epoch: 30, Loss 0.000
Epoch: 31, Loss 0.000
Epoch: 32, Loss 0.000
Epoch: 33, Loss 0.000
Epoch: 34, Loss 0.000
Epoch: 35, Loss 0.000
Epoch: 36, Loss 0.000
Epoch: 37, Loss 0.000
Epoch: 38, Loss 0.000
Epoch: 39, Loss 0.000
Epoch: 40, Loss 0.000
Epoch: 41, Loss 0.000
Epoch: 42, Loss 0.000
Epoch: 43, Loss 0.000
Epoch: 44, Loss 0.000
Epoch: 45, Loss 0.000
Epoch: 46, Loss 0.000
Epoch: 47, Loss 0.000
Epoch: 48, Loss 0.000
Epoch: 49, Loss 0.000
Epoch: 50, Loss 0.000
trained params
w: [0.05073234817652038, 0.007306286342947243, 0.023218625946243507]
b: 0.016648404245261664可以看到结果还是不错的
