GSoC Work Product of Yen-Chen Lin

GSoC_work_product.md

Merged Works

• #6593 Use fused types in sparse mean variance functions
• #6664 Make cython compiler directives global
• #6785 Make csr row norms support fused types
• #6846 Allows KMeans/MiniBatchKMeans to use float32 internally by using cython fused types
• #7053 Reduce duplicated code in _k_means.pyx
• #7170 Make Silhouette_score support sparse X
• #7184 Add whats_new about fixing example
• #7187 Rename CV params n_{folds,iter} to n_splits
• #6586 Add homogeneous time series cross validation
• #6913 Make CD use fused types
• #7218 Make enet_coordinate_descent_gram support fused types

Works in Progress

• #6540 Add ensemble selection algorithm
  • Implement main algorithm
  • Add example code
  • Determine API
  • Add tests
  • Detail documentation
• Make Neighbor Tree support Cython fused types
  • Find workaround to use void* attributes as fused types variables
  • Make code support fused types
  • Add tests
• Make multitask enet support Cython fused types
  • Make code support fused types
  • Add tests

Blog Posts

• WIP - Time Series Cross Validator
• 2016/8/8 - Workaround to use fused types class attributes
• 2016/7/23 - Interactive Cython with IPython, no compilation anymore!
• 2016/7/20 - How to set up 32bit scikit-learn on Mac without additional installation
• 2016/7/17 - Using Function Pointer to Maximize Code Reusability in Cython
• 2016/6/27 - scikit-learn KMeans Now Support Fused Types
• 2016/6/8 - Difference between np.float64 & np.float64_t
• 2016/5/27 - scikit-learn Sparse Utils Now Support Fused Types
• 2016/5/22 - Hello Google Summer of Code!

Development Log

See here for my detailed day by day development progress.

Cython Fused Types Limitation

• Doesn't work with class inheritance. See bug report here.
• Not supported as attributes of extension types.
• When having fused type arguments in extension class method, you cannot have default value arguments in the same method. See bug report here.
• Can be used to declare local variable of funtion if and only if any argument of function is of that fused type.

State of other fused types projects

• Sequential Dataset (SAG, SGD) - It seems to involve inheritance, which is known to not work with Cython fused types. See detailed analysis here.
• Neighbor Tree - I've found a workaround to use fused types class attributes (see this blog post), but still not completely make binary_tree.pxi support fused types. Both BallTree and KDTree can be modified to support fused types after that.

Misc

Some pharases I learned from my mentor:

• Lead someone down the garden path: Deceive someone.
• Whatever floats your boat: Do whatever you like.
• Getting things out the door: Getting things signed, sealed, delivered.

Usually one leads someone up the garden path. And I suggest you be cautious about using "whatever floats your boat"; it's very casual, and sometimes dismissive. :) I should probably look through the rest of this...

Regarding

When having fused type arguments in extension class method, you cannot have default value arguments in the same method. See bug example here.

Has this been reported to Cython? Please do so if it has not been. Worst they can do is clarify your error or identify it as a duplicate.

You can then emphasise that this is also a contribution of yours to the Python community.

I would appreciate a bit more detail on what held up "but still not completely make binary_tree.pxi support fused types." or whether it was just paused due to frustration.

I've submitted the bug of Cython to its Github repo, see issue here.