Fred, I think on_used_input=ignore should be a default in Matthew's theano printing code or that arg needs to be pushed up to his layer. I hit that issue too.
The code I have is here: https://github.com/PythonDynamics/pydy-code-gen See the results.txt file for basic speed comparisons. I'm generating the ODEs for an n-link pendulum with mechanics and then see how fast it generates and simulates with scipy.odeint. The code doesn't work at the moment. I haven't touched it in a month and looks like some things have changed in sympy and/or theano. I'll work on the bugs now. But "python benchmark.py" should run it with sympy master and ?some? version of Theano. Jason moorepants.info +01 530-601-9791 On Tue, Aug 6, 2013 at 9:25 AM, Frédéric Bastien <[email protected]> wrote: > Hi, > > I don't know what is sympy.functions.elementary.piecewise. Do Jason > answered that part? If not, I'll look into it to know how to make Theano > reproduce it. About converting any Sympy symbol to Theano symbol, when > there isn't a one to one matching, you can create a one to a full Theano > graph conversion. When this is possible, it is probably the best, as if you > make a new Theano op, it work work on the GPU. But if you make a Theano > graph, there is good change that the graph will already work on the GPU. > > If it is not possible to make a Theano graph for a sympy symbol, it is > possible to make a new Theano op that just wrap the sympy implementation. > Also, if this is a bottleneck, I recently added an example that show how to > use numba with Theano so speed up the python code in a Theano op. > > > Theano do not parallelize on the CPU, except for the call to BLAS, when > the BLAS library is parallel. On the GPU, it is parallel. > > Having on_unused_input=ignore is normal for complicated generated code. > That is why it was added. But when the code is simpler and not generated, > but all user manually coded, most of time it mean the user didn't do what > he wanted. If you know it is normal that you have unused input, there is no > problem to use that flag. > > > Jason, about the case where Theano is slower, can you send me the Theano > code? I would like to look at it. I'm very surprised that Theano is slower > then Sympy in this case and would like to know why it is like this. > > Fred > > > On Mon, Aug 5, 2013 at 9:03 PM, Jason Moore <[email protected]> wrote: > >> Guy, >> >> We're working on the same problem for sympy.physics.mechanics. Matthew >> Rocklin added support for matrix conversions in the theano code that is in >> SymPy and I used that, but found that theano was slower that lambdify for >> most of my cases (I only have two cores, so I'm not taking advantage of the >> Theano parallel stuff). I think writing specific code gen for ode >> integration is going to be the best bet. I'm happy to collaborate on this >> with you. >> >> >> Jason >> moorepants.info >> +01 530-601-9791 >> >> >> On Mon, Aug 5, 2013 at 3:13 PM, Guy Parsey <[email protected]> wrote: >> >>> Hello Everyone, >>> Thank you in advance for reading through my problem and for any input >>> you may have. I must say that I still feel like a novice programmer and my >>> problems may be easily solvable from a different mindset. My present >>> project entails time-integration of extremely stiff and non-linear ODEs >>> with regards to chemical kinetics (one derivative equation for each >>> variable species) and energy equations. Initially we were planning on using >>> the sympy.lambdify function to create callable functions for the main >>> function along with the jacobian and passing said functions to >>> scipy.integrate.odeint, however this method only works for easier test >>> cases (fewer species and/or no energy equations) before being limited by >>> either the list recursion limit or segfaulting due to the limited stack >>> size. I know that both of these limits can be edited, but that fact that I >>> am reaching them makes me feel as though I am doing something extremely >>> inefficiently. Outside of the documentation of SymPy and Theano, I have >>> also been heavily using the BlogPost by Matthew Rocklin >>> http://matthewrocklin.com/blog/work/2013/03/19/SymPy-Theano-part-1/ . >>> >>> Presently I am trying to use the mapping between Theano and SymPy >>> (sympy.printing.theanocode theano_function) to make my callable functions >>> and take advantage of the optimization routines. I have two major problems >>> and a few questions: >>> >>> 1st major problem: Though piecewise functions exist in SymPy >>> (sympy.functions.elementary.piecewise) there is no counterpart in Theano. >>> Looking at the source of the inspiration for theanocode ( >>> https://github.com/nouiz/theano_sympy/ graph_translation.py) I see >>> that some of the SymPy equivalents were defined as lambda functions. Is >>> there an equivalent way to add Theano conditional expressions wrapped into >>> a function to add to the mapping dictionary in theanocode.py? >>> >>> 2nd major problem: Similar to the problem above in that I am not sure >>> that the Theano counterpart is; some of the terms that I use are >>> interpolated functions (with one ODE variable as input) that we have >>> wrapped symbolically while providing a numerical implementation (so that >>> symbolic derivatives can be made, resulting in their own interpolations). >>> Is it possible to recreate the interpolation function as a Theano operation >>> for use within the system of ODEs? >>> >>> Remain questions: >>> I presently have to flatten my input to theano_function to a list of >>> expressions and then wrap to return to a form (Jacobian is a matrix not a >>> vector); is it possible to have a matrix of different expressions as an >>> input to theano_function with a vector output? >>> >>> I know that a huge amount of Theano speed up is due to parallelization >>> of matrix operations (which I do not have), should I be focusing on SymPy >>> Autowrap/Ufuncify or my own code generation instead of trying to get Theano >>> to play nicely? >>> >>> Stupid questions: >>> Does sympy.printing.theanocode.theano_function automatically optimize >>> the compiled graph? >>> >>> Minor comment: >>> Perhaps unnecessary for most uses of the theano_function, but I needed >>> to modify function inputs so as to be able to use the keyword argument >>> 'on_unused_input=ignore' as opposed to 'raise' so that I did not need to >>> have all symbols in all equations. This may be avoided by having the unused >>> symbols somehow (I don't know how) included in each expression. >>> >>> Thank you again for your time in reading my problems and any potential >>> help you may think of. I can attach code if necessary, I just didn't want >>> to make my post more confusing. >>> Have an excellent day. >>> Sincerely, >>> Guy Parsey >>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "sympy" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> To post to this group, send email to [email protected]. >>> Visit this group at http://groups.google.com/group/sympy. >>> For more options, visit https://groups.google.com/groups/opt_out. >>> >>> >>> >> >> -- >> You received this message because you are subscribed to the Google Groups >> "sympy" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To post to this group, send email to [email protected]. >> Visit this group at http://groups.google.com/group/sympy. >> For more options, visit https://groups.google.com/groups/opt_out. >> >> >> > > -- > You received this message because you are subscribed to the Google Groups > "sympy" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at http://groups.google.com/group/sympy. > For more options, visit https://groups.google.com/groups/opt_out. > > > -- You received this message because you are subscribed to the Google Groups "sympy" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/sympy. For more options, visit https://groups.google.com/groups/opt_out.
