RuntimeError: Expected scalar type Float but found Double

# RuntimeError: Expected scalar type Float but found Double

The PyTorch "RuntimeError: Expected scalar type Float but found Double" occurs when you pass an object of Double type when a Float tensor is expected.

To solve the error, use the float() method in the call to model().

Here is an example of how the error occurs.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()
input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()

print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

# ⛔️ RuntimeError: expected scalar type Double but found Float
# ⛔️ RuntimeError: expected scalar type Float but found Double
out = model(A)
print(out.data)

runtime error expected scalar type float but found double

The model() function expects to get called with a floating-point value but gets called with a double value.

# Convert the tensor to a float when calling `model()`

One way to solve the error is to convert the tensor to a float when calling the model() object.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()
input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()

print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

# ✅ Call float() method in call to model()
out = model(A.float()) # 👈️
print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.2995, 0.2112, 0.1614]], dtype=torch.float64)
tensor([[[[0.2408, 0.1265, 0.3745, 0.3934, 0.8293, 0.0694]]]],
       dtype=torch.float64)
tensor([[[[0.1593, 0.1805, 0.3291, 0.3042]]]])

use float method in call to model

The Tensor.float() method is equivalent to self.to(torch.float32).

Therefore, the following code sample is equivalent.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()
input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()

print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

# ✅ Convert the tensor to float
out = model(A.to(torch.float32)) # 👈️

print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.8885, 0.5696, 0.5680]], dtype=torch.float64)
tensor([[[[0.7581, 0.0192, 0.6453, 0.4124, 0.6039, 0.8607]]]],
       dtype=torch.float64)
tensor([[[[1.0510, 0.6189, 1.1513, 1.1993]]]])

using to method to convert tensor to float

Note that torch.float32 is Float and torch.float64 is Double.

# Call the `double()` method on the model

Your error message will contain the type that you need to convert your tensor or model to.

For example:

RuntimeError: expected scalar type Double but found Float
- Means that you need to convert the model to Double by calling the double(). function.
RuntimeError: Expected scalar type Float but found Double
- Means that you need to convert the model to Float by calling the float() function.

Here is an example of getting the first error.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()
input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()

print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

# ⛔️ RuntimeError: expected scalar type Double but found Float
out = model(A)

print(out.data)

Notice that the error message states: "expected scalar type Double", therefore, we have to call the double() method on the model.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()
input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()

print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
).double() # 👈️ call double() function here

model[0].weight.data.copy_(kernel)
A = Variable(input_)

out = model(A)

print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.8403, 0.9668, 0.8745]], dtype=torch.float64)
tensor([[[[0.2680, 0.4733, 0.3304, 0.9919, 0.6566, 0.7897]]]],
       dtype=torch.float64)
tensor([[[[0.9717, 1.5846, 1.8108, 2.1589]]]], dtype=torch.float64)

calling double method on the model

Depending on your error message, you might also have to call the float() or long() methods on the model.

# Calling `float()` directly on the input

You can also solve the error by calling float() directly on the input.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()


# 👇️ Call float() directly on the input
input_ = torch.from_numpy(
  np.random.random_sample((1, 1, 1, 6))
).float()


print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

out = model(A)

print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.1354, 0.4203, 0.9403]], dtype=torch.float64)
tensor([[[[0.3909, 0.9533, 0.5753, 0.9366, 0.6232, 0.6240]]]])
tensor([[[[0.9945, 1.2515, 1.0575, 0.9755]]]])

# Calling the `float()` method after instantiating the `Variable()` class

You can also solve the error by calling the float() method after instantiating the Variable() class.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()


input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()


print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)

# 👇️ Calling float() method here
A = Variable(input_).float()

out = model(A)

print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.3257, 0.1915, 0.0070]], dtype=torch.float64)
tensor([[[[0.3627, 0.7398, 0.4000, 0.6038, 0.8281, 0.2235]]]],
       dtype=torch.float64)
tensor([[[[0.2626, 0.3218, 0.2517, 0.3568]]]])

# Using `torch.set_default_dtype()` to solve the error

You can also use the torch.set_default_dtype() method to solve the error.

Here is an example.

main.py

Copied!
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable

# ✅ Set default dtype here
torch.set_default_dtype(torch.float64)

kernel = torch.from_numpy(np.random.random_sample((1, 3))).double()


input_ = torch.from_numpy(np.random.random_sample((1, 1, 1, 6))).double()


print(kernel)
print(input_)

model = nn.Sequential(
    nn.Conv2d(1, 1,
              (1, 3),
              stride=1,
              padding=0,
              bias=False)
)

model[0].weight.data.copy_(kernel)
A = Variable(input_)

out = model(A)

print(out.data)

The code for this article is available on GitHub

Running the code sample produces the following output.

shell

Copied!
tensor([[0.1479, 0.4833, 0.8837]])
tensor([[[[0.7442, 0.2410, 0.6052, 0.1779, 0.0842, 0.3530]]]])
tensor([[[[0.7614, 0.4853, 0.2499, 0.3789]]]])

Note that torch.float32 is Float and torch.float64 is Double.

So if you get an error, such as:

RuntimeError: expected scalar type Double but found Float

You would have to pass torch.float64 to the torch.set_default_dtype method

And if you get an error such as:

"RuntimeError: Expected scalar type Float but found Double"

You would have to pass torch.float32 to the method.

The torch.set_default_dtype() method sets the default floating-point dtype to the supplied type.

The method supports torch.float32 and torch.float64 as inputs.

# Additional Resources

You can learn more about the related topics by checking out the following tutorials:

I wrote a book in which I share everything I know about how to become a better, more efficient programmer.

You can use the search field on my Home Page to filter through all of my articles.

RuntimeError: Expected scalar type Float but found Double