Pytorch中一些重要的经典操作和简单讲解：
形状变换操作
reshape() / view()
python
import torch
x = torch.randn(2, 3, 4)
print(f"原始形状: {x.shape}")
# reshape可以处理非连续张量
y = x.reshape(6, 4)
print(f"reshape后: {y.shape}")
# view要求张量在内存中连续
z = x.view(2, 12)
print(f"view后: {z.shape}")transpose() / permute()
python
# transpose交换两个维度
x = torch.randn(2, 3, 4)
y = x.transpose(0, 2) # 交换第0和第2维
print(f"transpose后: {y.shape}") # torch.Size([4, 3, 2])
# permute重新排列所有维度
z = x.permute(2, 0, 1) # 将维度重排为 (4, 2, 3)
print(f"permute后: {z.shape}")拼接和分割操作
cat() / stack()
python
# cat在现有维度上拼接
x1 = torch.randn(2, 3)
x2 = torch.randn(2, 3)
# 在第0维拼接
cat_dim0 = torch.cat([x1, x2], dim=0) # (4, 3)
# 在第1维拼接
cat_dim1 = torch.cat([x1, x2], dim=1) # (2, 6)
# stack创建新维度并拼接
stacked = torch.stack([x1, x2], dim=0) # (2, 2, 3)chunk() / split()
python
x = torch.randn(6, 4)
# chunk均匀分割
chunks = torch.chunk(x, 3, dim=0) # 分成3块，每块(2, 4)
# split按指定大小分割
splits = torch.split(x, 2, dim=0) # 每块大小为2
splits_uneven = torch.split(x, [1, 2, 3], dim=0) # 不均匀分割索引和选择操作
gather() / scatter()
python
# gather根据索引收集元素
x = torch.randn(3, 4)
indices = torch.tensor([[0, 1], [2, 3], [1, 0]])
gathered = torch.gather(x, 1, indices) # (3, 2)
# scatter根据索引分散元素
src = torch.randn(3, 2)
scattered = torch.zeros(3, 4).scatter_(1, indices, src)masked_select() / where()
python
x = torch.randn(3, 4)
mask = x > 0
# 选择满足条件的元素
selected = torch.masked_select(x, mask)
# 条件选择
y = torch.randn(3, 4)
result = torch.where(mask, x, y) # mask为True选x，否则选y数学运算操作
clamp() / clip()
python
x = torch.randn(3, 4)
# 限制数值范围
clamped = torch.clamp(x, min=-1, max=1)
# 等价于
clipped = torch.clip(x, -1, 1)norm() / normalize()
python
x = torch.randn(3, 4)
# 计算范数
l2_norm = torch.norm(x, p=2, dim=1) # L2范数
l1_norm = torch.norm(x, p=1, dim=1) # L1范数
# 归一化
normalized = torch.nn.functional.normalize(x, p=2, dim=1)统计运算操作
mean() / sum() / std()
python
x = torch.randn(3, 4, 5)
# 各种统计量
mean_all = x.mean() # 全局均值
mean_dim = x.mean(dim=1) # 沿第1维求均值
sum_keepdim = x.sum(dim=1, keepdim=True) # 保持维度
# 最值操作
max_val, max_idx = torch.max(x, dim=1)
min_val, min_idx = torch.min(x, dim=1)广播和重复操作
expand() / repeat()
python
x = torch.randn(1, 3)
# expand不复制数据，只是改变视图
expanded = x.expand(4, 3) # (4, 3)
# repeat实际复制数据
repeated = x.repeat(4, 2) # (4, 6)tile() / repeat_interleave()
python
x = torch.tensor([1, 2, 3])
# tile像numpy的tile
tiled = x.tile(2, 3) # 重复2次每行，3次每列
# repeat_interleave每个元素重复
interleaved = x.repeat_interleave(2) # [1, 1, 2, 2, 3, 3]类型转换操作
to() / type() / cast()
python
x = torch.randn(3, 4)
# 类型转换
x_int = x.to(torch.int32)
x_float = x.type(torch.float64)
x_cuda = x.to('cuda') # 移到GPU（如果可用）
# 设备转换
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
x_device = x.to(device)在深度学习领域，这类张量运算操作具有极高的应用频率，尤其在数据预处理、模型架构构建及推理后处理等关键环节中不可或缺。熟练掌握此类算子的应用逻辑，能够显著优化张量数据的处理流程，提升深度学习任务的执行效率与工程实现效能。