Tensor new specification for scalar values

Hello,

I propose the following specification for handling scalar values with Tensor: https://gitlab.eclipse.org/groups/eclipse/aidge/-/wikis/Tensor#scalar-value

The TLDR is:

---

In order to handle numpy scalar values, we allow a tensor to have a no dimension (nb_dims==0) but a size of 1.

In the case of a Tensor with no values (nullptr), we need to have nbdims==0 and size==0.

This means that a simple check on the number of dimension is not enough to verify if an element is empty or not !

This behavior is required to have the best interroperability with numpy array:

>>> import numpy as np
>>> a = np.array(5)
>>> b = np.array([4])
>>> a.size
1
>>> a.ndim
0
>>> a.shape
()
>>> b.size
1
>>> a.ndim
0
>>> b.shape
(1,)

---

Also see Python documentation:

• Py Buffer protocol shape: https://docs.python.org/3/c-api/buffer.html#c.Py_buffer.shape
• Numpy Style array: https://docs.python.org/3/c-api/buffer.html#numpy-style-shape-and-strides

Can you @eclipse/aidge please tell me if you agree with this proposition ?

added Discussion Documentation Important labels

assigned to @cmoineau

changed the description

d01920a4 -> Introduce the proposed changes

PS: thx @pineapple for the help on checking what to modify

Makes sense to me to define scalar as 0-dim but beware of:

• getter/setter behavior, especially if using coordinates (what are coordinates in 0-dim, actually coordinates are implemented with std::vectors)
• consistency when computing/updating mSize
• what is mSizeM1 value (if it still exists)?
• what is an empty vector and how to check that (imho, direct testing of a member variable value, for that or anything else (see getImpl() return used to test for implem leading to issues now that it can throw), I prefer explicit boolean function such as hasImpl(), isEmpty(),...)

• I don't know what mSizeM1 refers to.
• With this specification, I think we forbid empty tensor, with the minimal allocation being one value.

This commit makes the change. !67 (d01920a4)

getter/setter behavior, especially if using coordinates (what are coordinates in 0-dim, actually coordinates are implemented with std::vectors)

Numpy library, in the very specific case of an array containing a scalar, uses a function item() to get the value from the numpy array object. This function only works for numpy scalars, we could do the same thing.

consistency when computing/updating mSize

The new function to handle mSize is consistent and faster than the previous one.

what is mSizeM1 value (if it still exists)?

This parameter has been removed in !57 (merged)

what is an empty vector and how to check that (imho, direct testing of a member variable value, for that or anything else (see getImpl() return used to test for implem leading to issues now that it can throw), I prefer explicit boolean function such as hasImpl(), isEmpty(),...)

I guess you meant an empty Tensor, we should set a new default Tensor.

closed

On second thought, not such a good idea, especially because of the confusion between a single element 1D tensor and a scalar. I think that Eigen and matlab(tm), for instance are using only tensor, event for scalars, at least for storage (Eigen proposes extended API for 1D, 2D and 3D vectors (which are 1D tensor of size 1,2 and 3). But if your confident on the way to manage that without introducing inconsistencies...