【PyTorch学习6】张量形状操作

• reshape

reshape函数用于改变数据的维度；

# 使用data.shape data.size()查看数据大小
# reshape 前后元素个数不变
data = torch.tensor([[1,2,3],[4,5,6]])  # torch.Size([2, 3])
data1 = data.reshape(3,2)  # torch.Size([3, 2])
# 使用-1省略形状
data2 = data.reshape(1,-1)  # torch.Size([1, 6])
data3 = data.reshape(-1,1)  # torch.Size([6, 1])

• transpose/permute

transpose把两个维度位置进行交换；permute一次性更换维度位置；

• view

只可用于存储在整块内存中的张量（有些张量由不同数据块组成（经过transpose/permute操作后使得张量不连续）——先使用contiguous函数使之在一块内存）；

# transpose 只可两个维度交换
data = torch.randint(0,10,[3,4,5])  # torch.Size([3, 4, 5])
data1 = data.transpose(0,2)  # torch.Size([5, 4, 3])  or data1 = torch.transpose(data,0,2)

# permute
data2 = data.permute(0,2,1)  # torch.Size([3, 5, 4])  or torch.permute(data,0,2,1)

# view
data3 = data.view(3,2,10)  # torch.Size([3, 2, 10])
print(data2.is_contiguous())  # False
data4 = data2.contiguous().view(3,10,2)  # torch.Size([3, 10, 2])

• squeeze/unsqueeze

squeeze删除shape为1的维度；unsqueeze在某一位置添加为1的维度；

data = torch.randint(0,10,[2,1,2,1])  # torch.Size([2, 1, 2])
# squeeze 默认去掉所有为1的维度
data1 = data.squeeze()  # torch.Size([2, 2])
# squeeze 指定去掉为1的维度
data2 = data.squeeze(1)  # torch.Size([2, 2, 1])

# unsqueeze
data3 = data.unsqueeze(0)  # torch.Size([1, 2, 1, 2, 1])
data4 = data.unsqueeze(-1)  # torch.Size([2, 1, 2, 1, 1])