jaberg / pyautodiff Goto Github PK

Track line numbers as Context steps through bytecode, for better error messages.

If Context evaluates a function that accepts **kwargs parameter, and then passes through those kwargs to another function, it doesn't work.

Support for this sort of thing:

a += 5
a[1, :] += 10

It would be good to ensure that this worked.

Theano nodes already have .name attribute, use it.

The batch costs are sums. The derivatives are sums. No need to compute all data set at once, it wastes huge memory. How to give interface to not do this?

Idea - give same streams interface to lbfgs as fmin_sgd has. Algorithm iterates over all blocks accumulating gradient between each call to L-BFGS.

The technique of shadowing temporaries by id was a good start, but not robust. Problems are:

• it has to disable gc (not a big deal)
• it is confused by a single object showing up twice by coincidence (especially small integers!)

What would make this better is a proper virtual machine class (with globals, locals, and a stack) that is instantiated once to run the original program, and once to run the symbolic shadow program.