Confusion about parameters of torch.addmm(...)

1

following is the definition of torch.addmm 's parameters:
torch. addmm ( beta=1 , mat , alpha=1 , mat1 , mat2 , out=None ) → Tensor
from torch — PyTorch 2.1 documentation
but under it there is an example:

M = torch.randn(2, 3)
mat1 = torch.randn(2, 3)
mat2 = torch.randn(3, 3)
torch.addmm(M, mat1, mat2)
tensor([[-4.8716, 1.4671, -1.3746],
[ 0.7573, -3.9555, -2.8681]])
the call of torch.addmm in the example only uses 3 parameters, when addmm’s last 2 parameters don’t heve a default value.
This is all my confusion about it.

2

What the M does is to add the random number generated by the 2 by 3 matrix while the mat1 and mat2 are multiply together
that
it will add mat1 and mat2 and multiply the final matrix by M
and the M rep the mat in the addmm

3

But what I see from pytorch doc is that " torch.addmm( beta=1 , mat , alpha=1 , mat1 , mat2 , out=None )" in which M，mat1 and mat2 should be seen as beta, mat and alpha. I run the example on my pc, it just works as you say, but I don’t understand why it works. Please cure my confusion. Thanks

4

alpha and beta are only scaling value in the vec matrix

5

aplha scale the value mul of mat1 and mat2 while beta scale the mat value

6

I know what torch.addmm does. But I don’t know how it can be like this.
let me take an example. why torch.addmm(1,M,1,mat1,mat2) and torch.addmm(M,mat1,mat2) can work the same? why the M in the second call can be parsed as mat instead of beta while the M in the second call is the first parameter？ Please forgive my English…

7

Because beta and alpha have positional argument and have a input value 1 from the function.

9

To my knowledge, the non-default parameter can’t follow the default parameter in python. And I’m also confused about the definition of addmm in pytoch doc, where the mat follows the beta=1. And this question is closely related to what we talk about.

10

sure,bt in that case alpha already have an input in the argument
let take for instance a function to calulate neural network

> def nn(weight,bias=1,features):
> 
>    return weight*feature+bias

so if you use the function note the bias in the function is given a default value 1 so it allow u to input weight and features without worring about bias except u want to change bias

12

xx
My python version is 3.6.6. My python don’t allow it. And does the ‘nn’ function you just define work on your device?

13

just using it to explian i did’t mean u should test

14

do u understand argument of a function very well

1 Like
15

What happens here is that the addmm does have “overloads” to implement the behaviour that, as you correctly note, would not be possible using a single plain Python function.
The twist is that the one using keyword only alpha and beta arguments is the preferred one (defined in aten/src/ATen/native/native_functions.yaml) while the others are deprecated (defined in tools/autograd/deprecated.yaml).
It is, perhaps, unfortunate that the addmm example uses parameters that are marked as deprecated.

Best regards

Thomas

(who looked at this way too much when generating typehints :slight_smile: )

1 Like
16

Other has explained to me that the definition isn’t what’s shown in pytorch doc. Any way, thanks

17

Though I can’t understand the content in the link, it’s enough for me to know the the definition isn’t what’s shown in pytorch doc. Thank you vvvvery much!