Hi Gerrit,
1. If I understand you correctly, the results for the 2D case do not
change. So I see no problem with the new implementation.
2. I don't think that there was a special reason to choose scipy.linalg.inv
over numpy.linalg.inv.
You are absolutely right, that the function(s) should be
On 22 August 2017 at 12:59, Oliver Lemke wrote:
> Hi Gerrit,
>
> Why is it necessary to replace .T with .swapaxes? Could you give an example
> were the former fails but the latter works?
K.T swaps indexes 0 and 1, but matmul broadcasts along indices -1 and
-2, so
Hi Gerrit,
Why is it necessary to replace .T with .swapaxes? Could you give an example
were the former fails but the latter works?
cheers,
/oliver
> On 22 Aug 2017, at 13:37, Gerrit Holl wrote:
>
> Hi,
>
> I need to adapt typhon.oem.retrieval_gain_matrix to allow
>
Hi,
I need to adapt typhon.oem.retrieval_gain_matrix to allow
broadcasting. That would require two changes:
- replace every occurrence of .T by .swapaxes(-1, -2)
- replace scipy.linalg.inv by numpy.linalg.inv, as the former does not
broadcast but the latter does.
The former is fully backward