-
Notifications
You must be signed in to change notification settings - Fork 32
Current work
This is the developers' to-do list. There are several sections, with earlier ones being higher priority than later ones. Tasks within each section are also ordered approximately by priority.
Most of our current tasks have been moved to GitHub's issue trackers; those are our primary record of outstanding tasks, so please check there first. In particular, new tasks should be submitted to GitHub rather than added to this list.
- Tasks to be addressed for the upcoming 1.0 release
- What the developers are working on most actively right now.
- Things we hope to take care of eventually
- Tasks of lower priority; if you would like to see one of these tasks completed any time soon, please volunteer (even if a developer is already assigned)!
- Ongoing work
- Long-term tasks that are underway, and tasks that the developers might currently be investigating: work with uncertain finishing times.
By each task, initials in parentheses typically indicate the main person working on an item, but others may also be involved. Items with no initials are not (yet) assigned to a specific developer (so please feel free to volunteer!!!!). Dates indicate when the item was first added to the list, or a change was made.
Priorities:
8: release 1.0
3: release someday
0: remove alert
* (8) boundingregion.py cleanup
* cf.py
(8) learning rate moves to learning function rather than cfprojection (CEB)
(8) where mask created (by cfprojection/cf) (CEB)
(8) learning rate a parameter of CFPLearningFn (CEB)
(8) CFPOutputFn mask parameter could be dropped now a masked iterator
can be passed in (CEB)
(8) JCALERT! We might want to change the default value of the
### input value to self.src.activity; but it fails, raising a
### type error. It probably has to be clarified why this is
### happening (CEB)
(3) calculation of no. of units (internal)
* functionfamily.py
(8) TransferFn: rename norm_value (CEB)
(8) LearningFn should have learning_rate param
(see same alert in cf.py) (CEB)
* projection.py
(8) other SheetMask + subclasses cleanup
* Slice
(8) M[slice]-style syntax (first figuring out performance implications
of attribute access) (CEB)
(8) general cleanup (CEB)
* PatternGenerator
(8) needs to support plasticity of output functions (after fixing
Pipeline's plasticity support)
* Simulation
(3) EPConnectionEvent always deepcopying data: does it need to?
(8) SomeTimer (also #1432101)
(8) the mess inside run(); how Forever etc is implemented
(3) PeriodicEventSequence
## JPHACKALERT: This should really be refactored into a
## PeriodicEvent class that periodically executes a single event,
## then the user can construct a periodic sequence using a
## combination of PeriodicEvent and EventSequence. This would
## change the behavior if the sequence length is longer than the
## period, but I'm not sure how important that is, and it might
## actually be useful the other way.
(8) gc alert in simulation.__new__ (CEB)
* Parameters
(3) script_repr:
# JABALERT: Only partially achieved so far -- objects of the same
# type and parameter values are treated as different, so anything
# for which instantiate == True is reported as being non-default.
(8) ParamOverrides should check_params() (CEB)
Add a section about plotting 'Orientation tuning fullfield' tuning curves. CB: would the tutorial benefit from being split up a little more? Maybe it's getting daunting?
- NOW A NOTEBOOK - OUTDATED?
no guarantee that code in/related to connectionfield is valid at all densities (tied to c++ comparisons task).
Once Kateryna's project completes, should systematically examine the example files (see above task) and eliminate as many as possible; several should now be able to be replicated using her master file. At the same time, should update most of the examples to use explicit parameters for v1aff_radius and lgnaff_radius, so that the sizes of the LGN and retina sheets can be calculated appropriately. Currently many of them have a size that includes a term 0.25 as a buffer, yet the relevant radius to be buffered against is 0.27083. There is no difference in the resulting matrix sizes at the default densities, but for high enough LGN densities we would expect that a few CFs around the edge would be cut off slightly using the current values.
(or wherever I tried to add test for not-run simulation before presenting patterns/saving generators)
test for recent run() problems - to catch problems with other future number types
- WHAT IS THIS ABOUT EXACTLY?
Need to implement more of the optimizations from the C++ LISSOM code.
Do we need our own simple timing functions to make it easier for users to optimize their components (as opposed to the overall Topographica framework, for which the current profile() commands are appropriate)? A facility for reporting the approximate time spent in methods of each EventProcessor? In any case, provide more guide for the user for doing optimization, focusing on the components we expect to be the bottlenecks. Add general advice for optimization to the manual pages.
Measurement of numpy.sum(X)/X.sum()/sum(X) performance. Difference between simulation results on different platforms (for slow-tests in Makefile).
defaults to something like lambda:0; the gui could set to topo.sim.time. Rather than having a default of topo.sim.time in places that don't know about simulations. In parameterclasses?
- OUTDATED - IMPLEMENTED AS ``param.Time``
- examples/joublin_bc96.ty
- examples/lissom_whisker_barrels.ty
- examples/ohzawa_science90.ty
And others...
We need a test with non-square input sheets, non-square LISSOM sheets, etc., with both types of non-squareness...and we also need to test whatever map measurement that we can (e.g. or maps). Could also add coverage testing, e.g. using figleaf or coverage.py.
- USING COVERALLS FOR PARAM UNIT TESTS
http://www.digitalpeers.com/pythondebugger/.
- OUTDATED - DEAD LINK. RECOMMEND IPDB.
Could consider using or taking components from: SciPy, ScientificPython, Chaco, Pyro (the robotics package), g, logger (instead of our custom messaging functions).
- OUTDATED - NO LONGER BUILDING EXTERNAL, USING PYTHON LOGGING
- numpy.__config__.show() - warn users if they're using a slow version? - http://www.scipy.org/Numpy_Example_List?highlight=%28example %29#head-c7a573f030ff7cbaea62baf219599b3976136bac >>> >>> import numpy >>> if id(dot) == id(numpy.core.multiarray.dot): # A way to know if you use fast blas/lapack or not. ... print "Not using blas/lapack!"
alterdot() numpy.alterdot(...) alterdot() changes all dot functions to use blas. (but note that dot() is, for some reason, not a dot product.)
Complete test file for Composite and Image. investigate failing test in testimage.py (that uses sheet functions). Currently commented out; may not be a problem.
Finish porting all categories of simulations from parts II and III of the LISSOM book (i.e. orientation maps, ocular dominance maps, direction maps, combined maps, face maps, and two-level maps) to Topographica.
Maybe should be one canvas with bitmaps drawn on. Then we'd get canvas methods (eg postscript()). But right-click code will need updating. Should be easy to lay out plots on a canvas, just like the grid() code that we have at the moment.
Might someday be interesting to have read-only objects, aiming at copy-on-write semantics, but this seems quite difficult to achieve in Python.
Optional packages (e.g. mlabwrap, pyaudio) on Windows and OS X.
- OUTDATED
And writing doc strings at the same time.
- TOO GENERAL!
Plus making sure base and simple classes are imported into packages (i.e. Sheet into topo/sheets/, Projection into topo/projections/, Constant into topo/patterns/, and so on).
Improving both, plus eliminating ALERTs. Could use Sphinx instead of epydoc for Reference Manual; apparently searchable.
- OUTDATED - NOW USING SPHINX
how to build scipy without requiring any of the external linear algebra libraries, etc? Then scipy would at least build easily, and users could install the optimized versions if they wished. Investigate garnumpy.
- OUTDATED - NO LONGER BUILDING SCIPY
Consider interfacing to this toolbox, which contains Matlab implementations of twenty techniques for dimensionality reduction. A number of these implementations were developed from scratch, whereas other implementations are based on software that is already available on the Web. http://www.cs.unimaas.nl/l.vandermaaten/dr CB: or consider a python/numpy alternative? E.g. http://mdp-toolkit.sourceforge.net/
- OUTDATED - FIRST LINK IS DEAD, SECOND LINK MENTIONED HERE
Consider interfacing to this toolbox, which is handy for generating novel 3D objects, e.g. to use as training stimuli (perhaps for somatosensory simulations?). http://www.psych.ndsu.nodak.edu/brady/downloads.html
- OUTDATED - DEAD LINK