How to Fix "View Size Is Not Compatible" in PyTorch

Q: How to fix view size is not compatible with input tensor's size and stride in PyTorch?

Use .reshape() instead of .view(), or call .contiguous().view(). This error happens when the tensor is not stored contiguously in memory, typically after operations like transpose(), permute(), or narrow(). The .reshape() method handles non-contiguous tensors automatically, while .view() requires contiguous memory layout. If you need a contiguous tensor specifically, use .contiguous() first.

Use .reshape() instead of .view(), or call .contiguous().view(). This happens when the tensor is not stored contiguously in memory after operations like transpose or permute.

The Error

RuntimeError: view size is not compatible with input tensor's size and stride
(at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

Fix 1: Replace .view() with .reshape() (Recommended)

# Instead of this:
x = x.transpose(1, 2).view(batch, -1)  # ERROR

# Do this:
x = x.transpose(1, 2).reshape(batch, -1)  # Works ✓

.reshape() handles both contiguous and non-contiguous tensors. It returns a view when possible and copies data when necessary.

Fix 2: Make Contiguous First

# If you need .view() specifically:
x = x.transpose(1, 2).contiguous().view(batch, -1)  # Works ✓

Why This Happens

These operations make tensors non-contiguous (they change how data is indexed without moving it in memory):

.transpose() / .t()
.permute()
.narrow()
.expand()
Slicing operations like x[:, ::2]

x = torch.randn(3, 4)
print(x.is_contiguous())              # True
print(x.t().is_contiguous())          # False
print(x.t().contiguous().is_contiguous())  # True

view() vs reshape()

.view() — requires contiguous memory, always returns a view (no data copy), fails on non-contiguous tensors
.reshape() — works on any tensor, returns a view if possible, copies data if needed

In most cases, .reshape() is the safer choice. Use .view() only when you specifically want to guarantee no data copy.