Chapter 10: Type Conversions and Casting¶

“Wrong dtype, wrong device — game over.”

🔄 10.1 Why Casting Matters¶

Tensor operations require matching types and devices. If not, you’ll face runtime errors like:

RuntimeError: expected scalar type Float but found Long

Or worse — silently incorrect results.

So casting properly is not just about avoiding bugs — it’s about making your models work as intended.

10.2 Basic Casting Methods¶

➤ Convert to Float, Long, Int, Bool¶

x = torch.tensor([1, 0, 1])
x.float()      # torch.float32
x.long()       # torch.int64
x.int()        # torch.int32
x.bool()       # torch.bool

These are shorthand wrappers around .to(dtype=...).

10.3 .to() — The Multipurpose Transformer¶

.to() can change: - The dtype - The device - Or both at once

x = x.to(torch.float32)
x = x.to('cuda')
x = x.to(torch.float16, device='cuda')

✅ Best practice: use .to() for cross-device + dtype-safe conversions in pipelines.

10.4 .type() — Less Flexible, Still Useful¶

x = x.type(torch.FloatTensor)

But it doesn’t support device changes, so .to() is preferred in most modern PyTorch code.

10.5 Matching Types in Ops¶

This error is extremely common when dealing with losses or metrics:

preds = torch.tensor([0.6, 0.2, 0.8])         # float32
labels = torch.tensor([1, 0, 1])              # int64
loss = torch.nn.BCELoss()
loss(preds, labels)   # ❌ Throws error

✅ Fix:¶

labels = labels.float()

Most loss functions expect both inputs to be float32, not integers.

10.6 One-Hot Encoding Pitfall¶

When using one-hot encodings, make sure your tensors are in float (or bool in some masking scenarios):

y = torch.tensor([0, 2])
one_hot = torch.nn.functional.one_hot(y, num_classes=3).float()

If you pass long-typed one-hot vectors into models, expect weird gradients or silent failures.

10.7 Converting Between Devices Safely¶

Avoid this:

x.cuda()

Use this:

x.to('cuda')  # Safer, cleaner, can be combined with dtype

To stay cross-platform, define your device once:

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
x = x.to(device)

🚫 10.8 Dangerous Casting Practices¶

Practice	Why It's Bad	Fix
Using `.data`	Bypasses autograd	Use `.detach()`
Forgetting `.float()`	Breaks loss functions	Always match model dtype
Manual `.cuda()` calls	Breaks portability	Use `.to(device)`
Mixing float64 + float32	Silently slows ops	Use consistent dtype

10.9 Summary¶

Use .float(), .long(), .bool() for quick casting.
Use .to(dtype, device) to control both type and placement.
Prefer .to() over .type() and never use .data.
Always check input/output types before applying loss or activation functions.
Define your device once and pass it around to keep your code portable.