Optimal .tolist() method
The most optimal and fastest way to convert PyTorch Tensor to Python List is to use the .tolist() method. It accepts a tensor, copies the data of that tensor, creates a list out of it, and returns it.
If the input is a 1D tensor, it returns a flat list. If an input is a 2D tensor, it returns a list of lists.
import torch # 1D Tensor tensor = torch.tensor([11, 21, 19]) print(tensor) # Output: tensor([11, 21, 19]) print(type(tensor)) # Output: <class 'torch.Tensor'> converted_list = tensor.tolist() print(converted_list) # Output: [11, 21, 19] print(type(converted_list)) # Output: <class 'list'>
In this code, we took a 1D tensor and tried to convert it into a list, printing their data types using the type() method.
2D Tensor to nested list
If your input is a 2D tensor, as I mentioned earlier, it will return a list of lists (nested list).
import torch # 2D Tensor tensor = torch.tensor([[11, 18, 19], [21, 48, 60]]) print(tensor) # Output: tensor([[11, 18, 19], # [21, 48, 60]]) print(type(tensor)) # Output: <class 'torch.Tensor'> converted_list = tensor.tolist() print(converted_list) # Output: [[11, 18, 19], [21, 48, 60]] print(type(converted_list)) # Output: <class 'list'>
GPU Tensor (CUDA)
Tensors mostly run the GPUs but Python lists can only run on CPUs. What if we want to convert GPU tensors to a list? Well, to do that, we need to first move the tensors from GPU to CPU using the .cpu() method. Then, we can perform the conversion using .tolist().
import torch # GPU Tensor (CUDA) tensor = torch.tensor([[11, 18, 19], [21, 48, 60]], device='cuda') converted_list = tensor.cpu().tolist() print(converted_list) # Output: [[11, 18, 19], [21, 48, 60]]
If you forget to move to the CPU, you will encounter an error.
Tensor with Gradients
What if the tensor has requires_grad=True? Well, in that case, we need to use the .detach() method. So tensor.detach().cpu().tolist() is the right approach to avoid autograd tracking.
import torch # Tensor with Gradients grad_tensor = torch.tensor([11., 19., 21.], requires_grad=True) converted_list = grad_tensor.detach().cpu().tolist() print(converted_list) # Output: [11.0, 19.0, 21.0]
Failing to detach gradients with .detach() retains autograd history.
0-dimensional tensor
If we work with A 0-dimensional tensor (scalar), the .tolist() method returns the Python scalar value. You can wrap it with “[]” to make it a list.
import torch # Scalar Tensors scalar_tensor = torch.tensor(11) converted_list = [scalar_tensor.tolist()] print(converted_list) # Output: [11]
Sparse Tensors
What if your input tensor is sparse? Well, first, you need to convert it into a dense tensor using the .to_dense() method and then use the .tolist() method.
import torch
# Sparse Tensors
sparse_tensor = torch.sparse_coo_tensor(indices=[[0, 1], [1, 2]],
values=[3, 4], size=(2, 3))
print(sparse_tensor)
# Output: tensor(indices=tensor([[0, 1],
# [1, 2]]),
# values=tensor([3, 4]),
# size=(2, 3), nnz=2, layout=torch.sparse_coo)
converted_list = sparse_tensor.to_dense().tolist()
print(converted_list)
# Output: [[0, 3, 0], [0, 0, 4]]
Non-Contiguous Tensors
If your input tensor is non-contiguous, use the .contiguous() method before the list conversion.
import torch # Non-Contiguous Tensors non_contiguous_tensor = torch.tensor([[1, 2], [3, 4]]).t() # Transposed (non-contiguous) contiguous_tensor = non_contiguous_tensor.contiguous() converted_list = contiguous_tensor.tolist() print(converted_list) # Output: [[1, 3], [2, 4]]
Using NumPy as an intermediate
The .numpy() method converts an input tensor into a numpy array and then chains the output with the .tolist() method to get the list. In this approach, we use a numpy array as an intermediate result, which you can use to perform some numeric operations and then convert to a list. This is not a direct approach.
import torch # Using NumPy as an Intermediate tensor = torch.tensor([[11, 18, 19], [21, 48, 60]]) converted_list = tensor.numpy().tolist() print(converted_list) # Output: [[11, 18, 19], [21, 48, 60]]
Only use the .numpy() approach if you want to execute numpy-related operations. Otherwise, use the .tolist() method directly on a tensor.
Using list comprehension
For fine-grained control, you can use list comprehension or for loop.1D Tensor
import torch tensor = torch.tensor([1, 2, 3]) main_list = [x.item() for x in tensor] print(main_list) # Output: [1, 2, 3]
2D Tensor
import torch # Using List Comprehension tensor = torch.tensor([[11, 18, 19], [21, 48, 60]]) converted_list = [item.tolist() for item in tensor] print(converted_list) # Output: [[11, 18, 19], [21, 48, 60]]If the input tensor is very large due to loop overhead, it is inefficient.
