II. Tensor Shape & Dimension Transforms (张量形状与维度变换)#

1. `Tensor.view()` / `Tensor.reshape()`#

Reshapes a Tensor without changing its data. view requires Contiguous Memory (连续内存); reshape handles non-contiguous cases automatically.

1
x = torch.arange(12)   # shape [12]
2
y = x.view(3, 4)       # shape [3, 4]
3
z = x.reshape(2, 6)    # shape [2, 6]
4
w = x.reshape(-1, 3)   # -1 auto-infers → shape [4, 3]

Note: Prefer reshape by default; switch to view only when you need to guarantee zero-copy memory sharing.

2. `torch.squeeze()` / `torch.unsqueeze()`#

squeeze: Removes dimensions of size 1. unsqueeze: Inserts a size-1 dimension at a specified position.

1
x = torch.zeros(1, 3, 1, 5)
2
y = x.squeeze()       # [3, 5]
3

4
z = torch.zeros(3, 5)
5
w = z.unsqueeze(0)    # [1, 3, 5]
6
v = z.unsqueeze(-1)   # [3, 5, 1]

Note: unsqueeze is frequently used for Broadcasting (广播): add a batch dimension (批次维度) or channel dimension (通道维度) to a vector.

3. `torch.cat()`#

Concatenates multiple Tensors along an existing dimension (已有维度). Does not create a new axis.

1
a = torch.zeros(2, 3)
2
b = torch.ones(4, 3)
3
c = torch.cat([a, b], dim=0)  # shape [6, 3]
4

5
d = torch.cat([torch.zeros(2, 2), torch.ones(2, 4)], dim=1)  # shape [2, 6]

Note: All Tensors must have the same shape on every dimension except the concatenation axis (拼接轴).

4. `torch.stack()`#

Stacks Tensors along a new dimension (新维度). All input Tensors must be exactly the same shape.

1
a = torch.tensor([1, 2, 3])
2
b = torch.tensor([4, 5, 6])
3
c = torch.stack([a, b])        # [2, 3]  — new dim=0
4
d = torch.stack([a, b], dim=1) # [3, 2]  — new dim=1

Note: Key difference from cat: stack requires identical shapes and always creates a new axis.

5. `Tensor.permute()`#

Reorders dimensions according to a specified axis order. Equivalent to NumPy's transpose(axes).

1
# Convert NCHW → NHWC
2
x = torch.zeros(8, 3, 224, 224)
3
y = x.permute(0, 2, 3, 1)  # shape [8, 224, 224, 3]

Note: After permute, the Tensor becomes Non-contiguous (非连续). Call .contiguous() if a subsequent op requires contiguous memory.

6. `Tensor.transpose()`#

Swaps exactly two specified dimensions. A simplified version of permute for axis swapping.

1
x = torch.zeros(4, 5, 6)
2
y = x.transpose(1, 2)   # shape [4, 6, 5]
3

4
m = torch.rand(3, 4)
5
mt = m.t()              # 2D matrix transpose → shape [4, 3]

Note: Use .transpose(-1, -2) for Batched Matrix Transpose (批量矩阵转置), valid over any batch dimension.

7. `torch.split()` / `torch.chunk()`#

split: Splits by specified sizes. chunk: Splits into equal pieces; the last chunk may be smaller.

1
x = torch.arange(10)
2
parts = torch.split(x, 3)   # (tensor([0,1,2]), tensor([3,4,5]), tensor([6,7,8]), tensor([9]))
3
chunks = torch.chunk(x, 3)  # 3 chunks: [0–3], [4–6], [7–9]

Note: Multi-GPU Sharding (多GPU分片) and DataLoader batch splitting internally rely on chunk / split logic.

8. `Tensor.flatten()`#

Flattens a Tensor to 1D, or flattens a specific range of dimensions.

1
x = torch.zeros(2, 3, 4)
2
y = x.flatten()       # shape [24]
3
z = x.flatten(1, 2)   # shape [2, 12]  — only flatten dims 1–2

Note: Most commonly used at the CNN → Fully Connected (全连接) transition. Equivalent to x.view(x.size(0), -1).

9. `torch.broadcast_to()`#

Broadcasts (广播) a Tensor to a target shape as a read-only view. No data is copied.

1
x = torch.tensor([1, 2, 3])      # shape [3]
2
y = torch.broadcast_to(x, (4, 3)) # shape [4, 3]

Note: Broadcasting is the underlying mechanism of most PyTorch arithmetic operations. Understanding it helps avoid Shape Errors (形状错误).

10. `Tensor.expand()`#

Expands size-1 dimensions to a specified size. Shares storage (no memory copy), unlike repeat.

1
x = torch.zeros(3, 1)
2
y = x.expand(3, 4)   # shape [3, 4] — zero memory copy
3
z = x.repeat(1, 4)   # shape [3, 4] — actual data copy

Note: expand results in a Non-contiguous Tensor; call .contiguous() or .clone() before writing.

💡 One-line Takeaway
Master reshape, squeeze/unsqueeze, cat/stack, and permute — together they cover 90% of all shape manipulation needs.

II. Tensor Shape & Dimension Transforms (张量形状与维度变换)#

1. Tensor.view() / Tensor.reshape()#

2. torch.squeeze() / torch.unsqueeze()#

3. torch.cat()#

4. torch.stack()#

5. Tensor.permute()#

6. Tensor.transpose()#

7. torch.split() / torch.chunk()#

8. Tensor.flatten()#

9. torch.broadcast_to()#

10. Tensor.expand()#