torch.as_tensor(): Converting a Valid Input into Tensor

The torch.as_tensor() method converts an input data, such as Python lists, NumPy arrays, or Scalars, into a PyTorch tensor while optionally sharing memory with the input data.

The above figure illustrates that the conversion from a list to a tensor is accomplished with a single method and a single argument.

The main difference between torch.as_tensor() and torch.tensor() is that the .as_tensor() method prioritizes memory efficiency by sharing data when possible, making it ideal for integrating external data into PyTorch workflows.

Syntax

torch.as_tensor(data, dtype=None, device=None)

Arguments

Name	Description
data (array_like)	It represents input data such as a tuple, a list, a numpy ndarray, a scalar, or other types.
dtype (torch.dtype, optional)	It is the desired data type of the returned tensor.
device (torch.device, optional)	It is the device used to create the tensor. By default, it is the device of data. However, you can specify an external device in this argument.

Converting a list to a tensor

The most common use case is when your input is a list and you want to convert it to a PyTorch tensor.

Let’s define a simple 1D list and perform the conversion.

import torch

original_list = [11, 2, 31, 4, 51]

print(original_list)
# Output: [11, 2, 31, 4, 51]

print(type(original_list))
# Output: <class 'list'>

list_to_tensor = torch.as_tensor(original_list)

print(list_to_tensor)
# Output: tensor([11,  2, 31,  4, 51])

print(type(list_to_tensor))
# Output: <class 'torch.Tensor'>

In the above code, with the help of as_tensor(), we easily converted the list to a tensor and verified the conversion using the built-in type() method.

Specifying a dtype

While converting data, we can specify the data type of the output tensor. Let’s take a look at the program below:

import torch

original_list = [11, 2, 31, 4, 51]

print(original_list)
# Output: [11, 2, 31, 4, 51]

print(type(original_list))
# Output: <class 'list'>

list_to_tensor_float = torch.as_tensor(original_list, dtype=torch.float64)

print(list_to_tensor_float)
# Output: tensor([11.,  2., 31.,  4., 51.], dtype=torch.float64)

print(type(list_to_tensor_float))
# Output: <class 'torch.Tensor'>

Specifying Device (CPU to GPU)

If a GPU is available, we can convert to Tensor on GPU using the device=’cuda’ argument.

import torch

# Input list
data = [11, 21, 19]

# Convert to tensor on GPU (if available)
if torch.cuda.is_available():
    tensor = torch.as_tensor(data, device='cuda')
    print(tensor)
else:
    print("CUDA not available")

If you run the above code on a CPU, it will produce the “CUDA not available” output.

Converting a numpy array with memory sharing

If an input is a numpy array, torch.as_tensor() shares the memory. Modifying the original numpy array affects the tensor, demonstrating zero-copy behavior.

import torch
import numpy as np

# Input NumPy array
np_array = np.array([[11.0, 2.0], [3.0, 41.0]])

# Convert to tensor
tensor = torch.as_tensor(np_array)
print(tensor)
# Output: tensor([[11., 2.], 
#                 [3., 41.]])

# Modify NumPy array
np_array[0, 0] = 99.0

print(tensor)
# Output: tensor([[99.,  2.], 
#                [3.,  41.]])

Handling scalars

You can convert a single scalar value to a 0-dimensional tensor.

import torch

# Input scalar
scalar = 21

# Convert to tensor
tensor = torch.as_tensor(scalar, dtype=torch.float32)

print(tensor)  
# Output: tensor(21.)

Incompatible data types

What if we try to convert a list of strings to a tensor? In this case, it will throw ValueError: too many dimensions ‘str’.

PyTorch does not support strings as elements of tensors. Since strings are not numeric, it throws an error.

import torch

# Input list with non-numeric data
string_of_list = ["a", "b", "c"]

try:
    tensor = torch.as_tensor(string_of_list)
except ValueError as e:
    print(f"ValueError: {e}")

# ValueError: too many dimensions 'str'

To avoid this type of error, ensure that your input data contains integer or float values.

 That’s all!