[GSoC 2015 Week 7&8]

In Week 7 and Week 8, I am mainly working on the optimization of local_fill_sink. This is more complicated than I thought. For details, check discussions here.

When the code of this PR is being used, the time costed by "canonicalize" is less than the original code. (12 passes in 166 seconds --> 10 passes in 155 seconds, tested with the user case on my computer). 

But these changes make theano fails on a test case, "test_local_mul_switch_sink". This test case is to check if the optimizer "local_mul_switch_sink" behaves correctly. Why does it fail? For short, in this test there is a fgraph like " (* (switch ...) (switch ...) )", if this optimizer is applied correctly, the mul op will sink under the switch op, so that expression like "(* value_x Nan)" can be avoided and end up with right result. 

What stops the optimizer is the assert node inserted into the graph. What I am working on now is to make MergeOptimizer deal with nodes with assert inputs. Of course this is already another optimization. For the failed test case, one way is to modify the code of local_mul_switch_sink, make it able to applied through assert nodes, but this is not a good way because it is not general.

Please reply here or send me email if you have any idea or comments. Thanks very much.

[GSoC 2015 Week 5&6]

In the Week 5 and Week 6, I was working on a new feature for debugging, to display the names of rejected optimizers when doing "replace_and_validate()". The PR is here. 

To implement this feature in one place in the code, the python library "inspect" is used. Anytime validation fails, the code will inspect the current stack frames and get to know which optimizer is the caller and whether there is "verbose" flag. Then the information for debugging can be displayed.

Besides, optimization for the local_fill_sink optimizer is also began here, the main idea on local_fill_sink is to make "fill" to sink in a recursive way to be more efficiency. 

For the inplace_elemwise optimizer, it is not merged because of the bad performance.