Hello,

It is nice to hear that Theano will be further improved. 

By the way,  I have created highly GPU-paralleled routines having to do 
with solving symmetric positive-definite matrix equations, 
and auto-regressive systems (Levinson). It also has gradients. .They are 
coded in C and compiled with CUDA, then linked into Theano using
the Magma interface. (basically, I coded in C, compiled using CUDA, linked 
them to libmagma.so, added the headers to magma.h
and some interface code to skcuda/magma.py. I also added Python class 
definitions in gpuarry/linalg.py and corresponding C interface
code in gpuarray/c_code ). I think it is very useful and fast code and 
fixes some deficiencies in Theano:  In Theano there is no fast and 
fully parallel GPU implementation for solving and/or getting the 
determinant of symmetric positive-definite systems.  
An example would be to compute the probability distribution of a 
multi-variate Gaussian distribution - given its symmetric
positive-definite covariance matrix - on a batch of data - very slow in 
Theano using scan, and not GPU parallel.   

Perhaps there is some way to merge them intoTheano, but I'd need help in 
that. I would suggest
adding them as a separate library, similar to magma. I could provide the 
code and the interface code.

Paul Baggenstoss



On Monday, August 24, 2020 at 6:25:51 PM UTC+2 Thomas wrote:

> Hi all,
>
> We, the PyMC development team have forked Theano and will continue to 
> develop it: https://github.com/pymc-devs/Theano-PyMC
>
> We have already done a lot of work in terms of changing the testing 
> framework to pytest, fixing stylistic issues, and applying black formatting.
>
> The most excited new development is adding a new JAX linker next to the 
> cython/c one: https://github.com/pymc-devs/Theano-PyMC/pull/21 
>
> This turns out to be quite straightforward. If you would like to help with 
> any aspect of this, we'd greatly appreciate it. Or if you have any PRs that 
> never got merged into Theano, please consider resubmitting them there.
>
> I think Theano is still an amazing framework that's mature, performant, 
> and has a very readable and hackable code-base.
>
> Best,
> Thomas
>

-- 

--- 
You received this message because you are subscribed to the Google Groups 
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/theano-users/335c0adc-62e0-4fbe-bc2d-a036ffc85162n%40googlegroups.com.

Reply via email to