🔢 Tensor
DocArray supports several tensor types that can you can use inside BaseDoc
.
The main ones are:
NdArray
for NumPy tensorsTorchTensor
for PyTorch tensorsTensorFlowTensor
for TensorFlow tensors
The three of them wrap their respective framework's tensor type.
Note
NdArray
and TorchTensor
are a subclass of their native tensor type. This means that they can be used natively in their framework.
Warning
TensorFlowTensor
stores the pure tf.Tensor
object inside the tensor
attribute. This is due to a limitation of the TensorFlow framework that prevents you from subclassing the tf.Tensor
object.
DocArray also supports AnyTensor
, which is the Union of the three previous tensor types.
This is a generic placeholder to specify that it can work with any tensor type (NumPy, PyTorch, TensorFlow).
Tensor Shape validation
All three tensor types support shape validation. This means that you can specify the shape of the tensor using type hint syntax: NdArray[100, 100]
, TorchTensor[100, 100]
, TensorFlowTensor[100, 100]
.
Let's take an example:
from docarray import BaseDoc
from docarray.typing import NdArray
class MyDoc(BaseDoc):
tensor: NdArray[100, 100]
If you try to pass a tensor with a different shape, an error will be raised:
1 validation error for MyDoc
tensor
cannot reshape array of size 20000 into shape (100,100) (type=value_error)
Whereas if you just pass a tensor with the correct shape, no error will be raised:
Axes validation
You can check that the number of axes is correct by specifying NdArray['x','y']
, TorchTensor['x','y']
, TensorFlowTensor['x','y']
.
from docarray import BaseDoc
from docarray.typing import NdArray
class MyDoc(BaseDoc):
tensor: NdArray['x', 'y']
Here you can only pass a tensor with two axes. np.zeros(10, 12)
will work, but np.zeros(10, 12, 3)
will raise an error.
Axis names
You can specify that two axes should have the same dimensions with the syntax NdArray['x', 'x']
, TorchTensor['x', 'x']
, TensorFlowTensor['x', 'x']
.
from docarray import BaseDoc
from docarray.typing import NdArray
class MyDoc(BaseDoc):
tensor: NdArray['x', 'x']
Here you can only pass a tensor with two axes that have the same dimensions. np.zeros(10, 10)
will work but np.zeros(10, 12)
will raise an error.
Arbitrary number of axis
To specify that your shape can have an arbitrary number of axes, use the syntax NdArray['x', ...]
, or NdArray[..., 'x']
.
from docarray import BaseDoc
from docarray.typing import NdArray
class MyDoc(BaseDoc):
tensor: NdArray[100, ...]
Here you can only pass a tensor with at least one axis with dimension 100. np.zeros(100, 10)
will work but np.zeros(10, 12)
will raise an error.
Tensor type validation
You don't need to directly instantiate the NdArray
, TorchTensor
, or TensorFlowTensor
by yourself.
Instead, you should use them as type hints on BaseDoc
fields, where they perform data validation.
During this process, BaseDoc
will cast the native tensor type into the respective DocArray tensor type.
Let's look at an example:
from docarray import BaseDoc
from docarray.typing import NdArray
import numpy as np
class MyDoc(BaseDoc):
tensor: NdArray
doc = MyDoc(tensor=np.zeros(100))
assert isinstance(doc.tensor, NdArray) # True
doc.tensor
is an NdArray
:
But since it inherits from np.ndarray
, you can also use it as a normal NumPy array. The same holds for PyTorch and TorchTensor
.
Type coercion with different tensor types
DocArray also supports type coercion between different tensor types. This mean that if you pass a different tensor type to a tensor field, it will be converted to the correct tensor type.
For instance, if you define a field of type TorchTensor
and you pass a NumPy array to it, it will be converted to a TorchTensor
.
from docarray import BaseDoc
from docarray.typing import TorchTensor
import numpy as np
class MyTensorsDoc(BaseDoc):
tensor: TorchTensor
doc = MyTensorsDoc(tensor=np.zeros(512))
doc.summary()
📄 MyTensorsDoc : 0a10f88 ...
╭─────────────────────┬────────────────────────────────────────────────────────╮
│ Attribute │ Value │
├─────────────────────┼────────────────────────────────────────────────────────┤
│ tensor: TorchTensor │ TorchTensor of shape (512,), dtype: torch.float64 │
╰─────────────────────┴────────────────────────────────────────────────────────╯
It also works in the other direction:
from docarray import BaseDoc
from docarray.typing import NdArray
import torch
class MyTensorsDoc(BaseDoc):
tensor: NdArray
doc = MyTensorsDoc(tensor=torch.zeros(512))
doc.summary()
📄 MyTensorsDoc : 157e6f5 ...
╭─────────────────┬────────────────────────────────────────────────────────────╮
│ Attribute │ Value │
├─────────────────┼────────────────────────────────────────────────────────────┤
│ tensor: NdArray │ NdArray of shape (512,), dtype: float32 │
╰─────────────────┴────────────────────────────────────────────────────────────╯
DocVec
with AnyTensor
DocVec
can be used with a BaseDoc
which has a field of AnyTensor
or any other Union of tensor types.
However, the DocVec
needs to know the tensor type of the tensor field beforehand to create the correct column.
You can specify these parameters with the tensor_type
parameter of the [DocVec
][docarray.vectorizer.doc_vec.DocVec] constructor:
from docarray import BaseDoc, DocVec
from docarray.typing import AnyTensor, NdArray
import numpy as np
class MyDoc(BaseDoc):
tensor: AnyTensor
docs = DocVec[MyDoc](
[MyDoc(tensor=np.zeros(100)) for _ in range(10)], tensor_type=NdArray
)
assert isinstance(docs.tensor, NdArray)
Note
NdArray
will be used by default if:
- you don't specify the
tensor_type
parameter - your tensor field is a Union of tensor or
AnyTensor
Compatibility of TorchTensor
and torch.compile()
PyTorch 2 introduced compilation support in the form of torch.compile()
.
Currently, torch.compile()
does not properly support subclasses of torch.Tensor
such as TorchTensor
.
The PyTorch team is currently working on a fix for this issue.
For a workaround to this issue, see the TorchTensor
API reference.