Description

igamenovoer

opened on Nov 14, 2023

Issue body actions

Initial Checks

• I have read and followed the docs and still think this is a bug

Description

when calling copy.deepcopy() over TorchTensor, it fails when tensor.dtype != torch.float32.
The problem is that TorchTensor.new_empty() will create a float32 tensor no matter what, regardless of self.dtype

Example Code

import torch
from docarray.typing.tensor.torch_tensor import TorchTensor
import copy

x = TorchTensor(torch.randn(2,3))
y = copy.deepcopy(x)    # this works

xx = TorchTensor(torch.randint(0,100,(2,3)))
yy = copy.deepcopy(xx)  # RuntimeError: Expected a Storage of type float or an UntypedStorage, but got type Long for argument 1 'storage'

Python, DocArray & OS Version

0.39.1

Affected Components

• Vector Database / Index
• Representing
• Sending
• storing
• multi modal data type

Activity

1. JoanFM commented on Nov 14, 2023

   

   JoanFM

   on Nov 14, 2023

   Member

   More actions

   Hey @igamenovoer,

   Thanms for reporting the issue. Since you spotted the problem, would you like to contribute the fix?
2. ai-naymul commented on Nov 30, 2023

   

   ai-naymul

   on Nov 30, 2023

   Contributor

   More actions

   I think we need to add a custom implementation of deepcopy for TorchTensor to avoid storage-sharing issues.

   @JoanFM could you please help me to contribute in this?

   Here is my approach:

   1. Create a deepcopy methods for that
   2. create a new tensor with the same data and properties
   3. Set the class to the custom Torchtensor class
   4. return the new tensor
3. JoanFM commented on Nov 30, 2023

   

   JoanFM

   on Nov 30, 2023

   Member

   More actions

   what do u mean by a deepcopy class for that?
4. ai-naymul commented on Nov 30, 2023

   

   ai-naymul

   on Nov 30, 2023

   Contributor

   More actions

   what do u mean by a deepcopy class for that?

   Oh, I am sorry I was mentioning the method, It's was a typing mistake.
5. JoanFM commented on Nov 30, 2023

   

   JoanFM

   on Nov 30, 2023

   Member

   More actions

   the approach seems to make sense indeed
6. ai-naymul commented on Nov 30, 2023

   

   ai-naymul

   on Nov 30, 2023

   Contributor

   More actions

   the approach seems to make sense indeed

   So should I try the approach?
7. JoanFM commented on Nov 30, 2023

   

   JoanFM

   on Nov 30, 2023

   Member

   More actions

   yws

   Reacted by Naymul Islam
8. ai-naymul commented on Dec 2, 2023

   

   ai-naymul

   on Dec 2, 2023

   Contributor

   More actions

   yws

   @JoanFM would you mind having a look at the PR, please?
9. JoanFM

   linked a pull request that will close this issuefix: fix storage issue in torchtensor class #1833on Dec 9, 2023
10. JoanFM

    closed this as completedin #1833on Dec 9, 2023

Issue actions