fix: enable torchembedding deepcopy - #1143

After the modification

import torch
from docarray import BaseDocument
from docarray.typing.tensor.embedding.torch import TorchEmbedding

class Meow(BaseDocument):
    embedding : TorchEmbedding

a = Meow(embedding=torch.randn(42))
print(id(a.embedding), id(a.copy(deep=True).embedding))

output has no error

140248435632272 140248435636032

Can u add to test to check that it is actual deep copied and not shallow copy ?

Can you explain this fix briefly? It is not obvious to me why this is fixing it

Can you explain this fix briefly? It is not obvious to me why this is fixing it

The deepcopy error is thrown from here https://github.com/pytorch/pytorch/blob/master/torch/_tensor.py#L170, and lines above are

new_tensor = self.new_empty([])
if type(new_tensor) is not type(self):

where type(new_tensor) is <class 'torch.Tensor'> and
type(self) is <class 'docarray.typing.tensor.embedding.torch.TorchEmbedding'>.

So according to the error msg, I should implement new_empty() for the subclass, which means the new_tensor should also be a TorchEmbedding.

Can you explain this fix briefly? It is not obvious to me why this is fixing it

The deepcopy error is thrown from here https://github.com/pytorch/pytorch/blob/master/torch/_tensor.py#L170, and lines above are

new_tensor = self.new_empty([])
if type(new_tensor) is not type(self):

where type(new_tensor) is <class 'torch.Tensor'> and type(self) is <class 'docarray.typing.tensor.embedding.torch.TorchEmbedding'>.

So according to the error msg, I should implement new_empty() for the subclass, which means the new_tensor should also be a TorchEmbedding.

@AnneYang720 can u add this to the docstring of new_empty in your PR so that ppl who will read the code will understand the same ?

📝 Docs are deployed on https://ft-fix-torchembedding-deepcopy--jina-docs.netlify.app 🎉