Features cumsum & cumprod #524
Conversation
heat/core/tests/test_arithmetics.py
Outdated
| cprod = ht.cumprod(a, 0) | ||
| self.assertTrue(ht.equal(cprod, result)) | ||
|
|
||
| a = ht.full((2, 4, 2), 2, dtype=ht.float32, split=1, device=ht_device) |
heat/core/arithmetics.py
Outdated
|
|
||
| recv = torch.ones(cprod.shape[:axis] + cprod.shape[axis + 1 :], dtype=send.dtype) | ||
|
|
||
| a.comm.Exscan(send, recv, MPI.PROD) |
There was a problem hiding this comment.
@krajsek does this work with the upcoming AD stuff?
heat/core/arithmetics.py
Outdated
| cumprod : DNDarray | ||
| A new array holding the result is returned unless `out` is | ||
| specified, in which case a reference to out is returned. | ||
| """ |
There was a problem hiding this comment.
Add Raises section for exceptions that are thrown from within this function
There was a problem hiding this comment.
Shall the raises section only be in __cum_op() or also here?
heat/core/arithmetics.py
Outdated
|
|
||
| recv = torch.ones(cprod.shape[:axis] + cprod.shape[axis + 1 :], dtype=send.dtype) | ||
|
|
||
| a.comm.Exscan(send, recv, MPI.PROD) |
heat/core/arithmetics.py
Outdated
| `axis` is not None or `a` is a 1-d array. | ||
|
|
||
| """ | ||
| if axis is None: |
heat/core/arithmetics.py
Outdated
|
|
||
| recv = torch.zeros(csum.shape[:axis] + csum.shape[axis + 1 :], dtype=send.dtype) | ||
|
|
||
| a.comm.Exscan(send, recv, MPI.SUM) |
There was a problem hiding this comment.
More or less identical code as above. Could you please introduce a __cum_op() call analogous to __reduce_op() in operations.py and make use of it?
heat/core/arithmetics.py
Outdated
| if csum.numel() > 0: | ||
| indices = -1 | ||
| Ni, Nk = csum.shape[:axis], csum.shape[axis + 1 :] | ||
| for ii in np.ndindex(Ni): |
There was a problem hiding this comment.
Can we somehow manage to do this without the Python loops via views? Performance on them should be slow.
There was a problem hiding this comment.
I replaced it with torch.index_select()
…-analytics/heat into enhancement/243-cum_operations
Codecov Report
@@ Coverage Diff @@
## master #524 +/- ##
==========================================
- Coverage 96.38% 96.35% -0.03%
==========================================
Files 75 75
Lines 14481 14585 +104
==========================================
+ Hits 13958 14054 +96
- Misses 523 531 +8
Continue to review full report at Codecov.
|
Description
Include a summary of the change/s.
Please also include relevant motivation and context. List any dependencies that are required for this change.
implementation of cumsum and cumprod functions. numpy's axis=None is not supported.
Issue/s resolved: #243
Changes proposed:
Type of change
Remove irrelevant options:
Due Diligence
Does this change modify the behaviour of other functions? If so, which?
yes / no