Test insertion of simplex with NaN filtration value - #424
Conversation
tlacombe
left a comment
tlacombe
left a comment
There was a problem hiding this comment.
From my (biased) perspective, it seems to be a nice addition. Perhaps it is almost useless (I don't really know how people use gudhi in general), but when you typically want to build a lower start filtration, starting from a complex + values on its vertices, this might simplify things a bit.
| st.insert([4], 5.) | ||
| st.insert([2], -1.) | ||
| st.insert([1, 2, 4], math.nan) | ||
| assert math.isnan(st.filtration([2, 4])) |
There was a problem hiding this comment.
| assert math.isnan(st.filtration([2, 4])) | |
| assert math.isnan(st.filtration([2, 4])) | |
| assert st.filtration([4]) == 5. | |
| assert st.filtration([2]) == -1. | |
| assert st.filtration([1]) == 2. | |
| assert st.filtration([1,2]) == 2. |
Perhaps we may want to make sure that we did not modify any of the previous values even before calling st.make_persistence_non_decreasing().
There was a problem hiding this comment.
I thought that checking after make_persistence_non_decreasing would be sufficient (if we overwrite a value with insert, make_persistence_non_decreasing won't be able to reinvent that value), but why not, done.
| st = SimplexTree() | ||
| st.insert([1], 2.) | ||
| st.insert([4], 5.) | ||
| st.insert([2], -1.) |
There was a problem hiding this comment.
| st.insert([2], -1.) | |
| st.insert([2], -1.) | |
| st.insert([1,2], 2.) |
Perhaps already add a single edge (with an admissible value) to make sure that both inserting nan and make_filtration_non_decreasing() do not play any role there
There was a problem hiding this comment.
I inserted an extra edge.
| .. note:: | ||
|
|
||
| Inserting a simplex with filtration value `math.nan` does not | ||
| modify the filtration value of any simplex already present. |
There was a problem hiding this comment.
Perhaps complete this note by adding a warning in the vein of what is done for assign_filtration() saying (e.g.) "In this particular case, the result will not be a valid filtration anymore. Callers are responsible for fixing this (using assign_filtration() or make_filtration_non_decreasing() for instance) before calling any function that relies on the filtration property, like persistence().
There was a problem hiding this comment.
I hadn't realized it, but having some NaN in the complex can have some weird effects. In particular, if I have vertex 0 with value NaN, inserting edge [0, 1] with value 42 does not change the value of vertex 0. I added some confusing text to the doc. We could change that behavior, but would need to check that it doesn't slow things down for normal operations.
|
It would be nice to test it at C++ level. Here is a gist that translates the python test in C++. |
3 participants
This patch documents (and tests) that inserting simplices with filtration value NaN does not modify the value for existing simplices.
Now to discuss whether that's actually the behavior we want 😉