Python打卡训练营学习记录Day48

作业：自己多借助ai举几个例子帮助自己理解即可

import torch
import numpy as np
 
# ================ 随机张量生成：torch.randn函数 ================
print("\n===== 随机张量生成示例 =====")
 
# 基础用法：生成指定形状的张量
tensor = torch.randn(3, 4)
print("形状为(3,4)的随机张量：\n", tensor)
 
# 结合维度扩展：生成高维张量
batch_tensor = torch.randn(2, 3, 4)
print("形状为(2,3,4)的随机张量（前两个元素）：\n", batch_tensor[:2])
 
# 与设备交互：指定张量存储位置
if torch.cuda.is_available():
    gpu_tensor = torch.randn(2, 2, device='cuda')
    print("GPU上的随机张量：\n", gpu_tensor)
 
# 生成指定均值和方差的张量
custom_tensor = 5 + 2 * torch.randn(3, 3)
print("自定义分布的随机张量（均值5，标准差2）：\n", custom_tensor)
 
# 生成固定随机数（用于实验可复现性）
torch.manual_seed(42)
fixed_tensor = torch.randn(2, 2)
print("固定随机种子的张量：\n", fixed_tensor)
 
 
# ================ PyTorch广播机制：加法与乘法 ================
print("\n===== 广播机制示例 =====")
 
# 示例1：一维张量加法
a = torch.tensor([1, 2, 3])  # 形状(3,)
b = torch.tensor([4])       # 形状(1,)
c = a + b                   # 等价于 a + [4, 4, 4]
print("一维加法广播结果：", c)  # 输出：tensor([5, 6, 7])
 
# 示例2：二维张量乘法（矩阵与向量）
matrix = torch.randn(3, 4)
vector = torch.randn(4)
result = matrix * vector  # 等价于矩阵每行与vector对应元素相乘
print("二维乘法广播结果形状：", result.shape)  # 输出：torch.Size([3, 4])
 
# 示例3：高维张量广播（三维+二维）
tensor1 = torch.randn(2, 1, 4)
tensor2 = torch.randn(3, 4)
result = tensor1 + tensor2  # 广播后形状：(2, 3, 4)
print("三维+二维广播结果形状：", result.shape)  # 输出：torch.Size([2, 3, 4])
 
# 多维张量与标量的运算
tensor = torch.randn(2, 3, 4)
result = tensor + 10  # 标量10自动广播为与tensor相同形状
print("张量+标量的结果形状：", result.shape)  # 输出：torch.Size([2, 3, 4])
 
# 广播机制在归一化中的应用
data = torch.randn(100, 5)
means = data.mean(dim=0)
stds = data.std(dim=0)
normalized_data = (data - means) / stds  # 自动广播(5,)为(100, 5)
print("标准化后数据形状：", normalized_data.shape)  # 输出：torch.Size([100, 5])
 
 
# ================ 广播机制与Numpy的一致性 ================
print("\n===== Numpy与PyTorch广播对比 =====")
 
# Numpy与PyTorch广播机制对比验证
np_a = np.random.randn(3, 1)  # 形状(3, 1)
np_b = np.random.randn(1, 4)  # 形状(1, 4)
np_result = np_a + np_b      # 结果形状(3, 4)
 
torch_a = torch.from_numpy(np_a)
torch_b = torch.from_numpy(np_b)
torch_result = torch_a + torch_b
 
print("Numpy结果形状：", np_result.shape)      # 输出：(3, 4)
print("PyTorch结果形状：", torch_result.shape)  # 输出：torch.Size([3, 4])
 
 
# ================ 广播机制错误案例分析 ================
print("\n===== 广播错误案例 =====")
 
# 错误案例：形状不兼容导致广播失败
tensor_a = torch.randn(3, 4)
tensor_b = torch.randn(2, 3)
 
try:
    result = tensor_a + tensor_b
except RuntimeError as e:
    print(f"广播错误：{e}")
    # 输出：The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 1    

@浙大疏锦行