Implement dask.array.broadcast_to #108
Conversation
| % (x.shape, shape)) | ||
|
|
||
| name = next(names) | ||
| blockdims = (tuple((s,) for s in shape[:ndim_new]) |
There was a problem hiding this comment.
Do we always want solid, un-cut blocks in the new dimensions? Or maybe this is a case where we reblock afterwards.
There was a problem hiding this comment.
I was thinking about this to. Arguably, this function should take optional blockshape and blockdims arguments.
I do think (auto-)reblocking will cover most of that use case. In particular, there is never a concern about memory usage because the new, arbitrarily large chunks will always be broadcast versions of the original chunks.
One potential downside of separate reblocking is that it wouldn't be able to take advantage of the fact that many of the chunks would be redundant. So that would entail somewhat more overhead, both for the scheduler and for slicing the broadcast expanded dimensions with numpy.
I'll make a follow-up issue.
|
Seeing other people implement dask functions gives me a very warm feeling inside. |
| x = np.random.randint(10, size=(5, 1, 6)) | ||
| a = from_array(x, blockshape=(3, 1, 3)) | ||
|
|
||
| for shape in [(5, 4, 6), (2, 5, 1, 6), (3, 4, 5, 4, 6)]: |
There was a problem hiding this comment.
Seeing these shapes helped me to understand the scope of broadcast_to
Implement dask.array.broadcast_to
No description provided.