How to Fix "Expected Scalar Type Float" in PyTorch

In PyTorch, this error means your input is float64 while the model weights are float32; fix it by calling x = x.float() before the forward pass.

Common Causes and Fixes

1. NumPy float64 to PyTorch (most common)

import numpy as np
import torch

data = np.array([1.0, 2.0, 3.0])    # float64 by default
tensor = torch.from_numpy(data)      # torch.float64 (Double)
# model(tensor) # ERROR: expected Float, got Double

# Fix:
tensor = torch.from_numpy(data).float()  # torch.float32 ✓
# or
tensor = torch.tensor(data, dtype=torch.float32) # ✓

2. Image loaded as uint8

# Raw image data is 0-255 uint8
image = torch.ByteTensor([[[128, 255], [0, 64]]]) # torch.uint8
# model(image) # ERROR: expected Float, got Byte

# Fix: convert and normalize
image = image.float() / 255.0  # torch.float32, range [0, 1] ✓

3. Mixing float32 model with float64 input

# Models default to float32
model = nn.Linear(10, 5)             # float32 weights
x = torch.randn(32, 10, dtype=torch.float64)   # float64 input
# model(x) # ERROR

# Fix: match the model's dtype
x = x.float()  # convert to float32 ✓
# or convert model to float64 (not recommended):
# model = model.double()

Check Dtype

print(x.dtype)          # torch.float64, torch.uint8, etc.
print(next(model.parameters()).dtype) # torch.float32 (typical)

Quick Reference

x.float()  # -> torch.float32 (most common)
x.double()  # -> torch.float64
x.half()   # -> torch.float16
x.int()   # -> torch.int32
x.long()   # -> torch.int64 (for labels/indices)