Could regmax also work on float arrays? about mahotas HOT 11 CLOSED

You are right. There is no reason for this function to not work with
float arrays.

from mahotas.

I think the same issue affects cwatershed (my goal right now is to simplify the "distance and watershed" example to avoid having to use stretch, which has issues when some distances are much larger than the relevant ones and thus cause the actually relevant distances to be squeezed into a single value -- yes, this can obviously be fixed by clipping the distances to a maximal relevant value, but if everything handled floats this would be a non-issue from the beginning).

from mahotas.

Yes, you are right on watershed again. The implementation currently has
integer hardcoded, but there is no reason to do it like that.

Adapting the code is actually fairly trivial. Are you available for
testing it?

from mahotas.

Not until the end of the month, but after that, yes.

from mahotas.

Hi, any news on that topic? I can help with testing now.

from mahotas.

Thanks for the ping. regmax/min now should work for FP images. Watershed is work-in-progress.

from mahotas.

I now added support for watershed too. I did some minimal (smoke) testing [included in unit tests], but I'd appreciate it if you gave it a thorough testing run.

from mahotas.

I looked at the current tests and the best I can think of is to change test_watershed and test_mix_types to also try float entries. Anything else you would suggest?
I also tried it on one of my actual programs and it seems to work fine, but that can hardly be turned into a unit test.

from mahotas.

If it seems to be working for your data, that's OK. If you do find any issue, just open a report. Thanks.

from mahotas.

Actually, the following seemingly reasonable patch to test_watershed leads to some failing tests. I haven't investigated further, though.

@@ -24,7 +24,7 @@ def test_watershed():
         [0,0,0,0],
         ])
     def cast_test(M,S,dtype):
-        M = M.astype(dtype)
+        M = M.astype(dtype if issubclass(dtype, (np.integer, int)) else int)
         S = S.astype(dtype)
         W = mahotas.cwatershed(2-S,M)
         assert sys.getrefcount(W) == 2
@@ -35,7 +35,8 @@ def test_watershed():
                [2, 2, 2, 2],
                [2, 2, 2, 2],
                [2, 2, 2, 2]]))
-    for d in [np.uint8, np.int8, np.uint16, np.int16, np.int32, np.uint32,int]:
+    for d in [np.uint8, np.int8, np.uint16, np.int16, np.int32, np.uint32, int,
+              np.float16, np.float32, np.float64, np.float128, float]:
         yield cast_test, M, S, d

from mahotas.

The markers argument should always be integer, but the other test is good and I can reproduce it.

from mahotas.