>From a hijacked thread in sympy-patches: ma., 05.07.2010 kl. 14.04 -0700, skrev Brian Granger: > On Mon, Jul 5, 2010 at 1:28 PM, Ronan Lamy <ronan.l...@gmail.com> wrote: > > Le lundi 05 juillet 2010 à 19:17 +0200, Øyvind Jensen a écrit : > >> > Yep. I just got back from SciPy and I talked a lot with one of the > >> > devs of theano: > >> > > >> > http://deeplearning.net/software/theano/ > >> > >> Thanks for bringing theano to my attention! > > > > Yes, that's very interesting! > > > >> > Are you familiar with this? It looks quite impressive and does code > >> > generation for both CPUs and GPUs. They actually have a lightweight > >> > symbolic engine and they were interested in seeing if they could use > >> > sympy instead. It is definitely worst looking at. I can contact them > >> > if you want to explore this further. > >> > > >> > >> Theano looks very nice, especially the GPU functionality is cool. I > >> will play with it and see what I can learn from it. However, they use > >> expression trees that are different from ours, so I fear it will be > >> nontrivial to use it directly with sympy. Whereas our tree consist of > >> operations and operands, they have Apply-nodes that contain an operation > >> instance and the operands. I wonder how much their optimization > >> framework depend on the form of the theano expression tree. > > > > Theirs is the obviously correct way of dealing with mathematical > > functions in a CAS, but sympy will get there in the end (that's issue > > 1688). In the mean time, there's an obvious bijection between f(x, y, z) > > and Apply(f, (x, y, z)) so this doesn't seem to be a difficult problem. > > On the other hand, theano uses types and variables, which sympy doesn't > > have: I think this is a more significant barrier to translating between > > the two. > > I talked to James, one of the Theano devs, and I think it wouldn't be > too difficult to use Sympy for the top level stuff, because of the > bijection between the two data structures. In terms of the types, my > idea was to use the assumptions system. > > > In any case, merging the two code bases seems a useful and achievable > > medium-term goal. > > Definitely. >
You are right. The bijection between the expression trees is obvious, and here[0] is an implementation. So far, there is only methods to convert the tree structure. So, essentially, I have implemented a mapping between two different forms of SymPy expression trees. So, how can the nodes be mapped to the right type in theano? I agree that assumptions can be part of the solution, in particular as a user interface to specify the datatype as float16, int32 etc. But, in order to manipulate array expressions correctly in sympy, I think it will be much easier to do it with something like Indexed [1], rather than symbols with elaborate assumptions. I am going to improve the Indexed module with functions to check conformance of arrays, and it could be an idea to implement checks of numpy broadcasting. In fact, I think now that numpy compatibility should be stated as an explicit goal of the Indexed class. Cheers, Øyvind [0]: http://github.com/jegerjensen/sympy/tree/theano [1]: http://github.com/jegerjensen/sympy/blob/fortran_codegen4/sympy/tensor/indexed.py > Cheers, > > Brian > > > Ronan > > > > -- > > You received this message because you are subscribed to the Google Groups > > "sympy-patches" group. > > To post to this group, send email to sympy-patc...@googlegroups.com. > > To unsubscribe from this group, send email to > > sympy-patches+unsubscr...@googlegroups.com. > > For more options, visit this group at > > http://groups.google.com/group/sympy-patches?hl=en. > > > > > > > > -- > Brian E. Granger, Ph.D. > Assistant Professor of Physics > Cal Poly State University, San Luis Obispo > bgran...@calpoly.edu > elliso...@gmail.com > -- You received this message because you are subscribed to the Google Groups "sympy" group. To post to this group, send email to sy...@googlegroups.com. To unsubscribe from this group, send email to sympy+unsubscr...@googlegroups.com. For more options, visit this group at http://groups.google.com/group/sympy?hl=en.
