DOC: See also np.load and np.memmap in np.lib.format.open_memmap

[f0k]

np.load supports memmap_mode to open existing .npy files as memory-mapped files.
np.memmap supports creating or loading raw binary files as memory-mapped files.
Neither support creating .npy files as memory-mapped files, which is handy when writing data that is too large to keep in memory at once while still retaining the advantages of the .npy format (shape/dtype information and easy loading). np.lib.format.open_memmap allows just this, but is difficult to find (even if you already know it exists!).

This adds a link to np.lib.format.open_memmap from np.load and np.memmap.

[charris]

Why do you want to advertise this function? It is already available in np.load and isn't imported into the numpy namespace. It looks more like a function that should be private. The documentation of load could use a bit of clarification for the mmap_mode parameter, noting that None is the default and reads the file into memory, as opposed to memmapping it.

[mhvk]

Hmm, I've just started using memmap files as I was trying to do things on my laptop when not having access to a bigger computer, and had turned to open_memmap to access .npy files, and I see that np.load just does it for me. Including, it seems, opening such files for writing, which it seems is what @f0k missed.

Given this, I agree open_memmap does not need to be advertised, but it would be helpful to put a note in memmap pointing to load as the way to map .npy files.

[mhvk]

Actually, the writing is problematic: how would I create a new array on disk with given dtype and shape, i.e., how could I use np.load to do the equivalent of (from somewhere in my code):

    d = open_memmap(out_fn, mode='w+', dtype=np.float16,
                    shape=(size, len(thread_ids)))

[f0k]

Why do you want to advertise this function? It is already available in np.load

No, this functionality isn't. You cannot create a new mem-mapped .npy file from scratch using np.load. Or can you?

Actually, the writing is problematic: how would I create a new array on disk with given dtype and shape, i.e., how could I use np.load to do the equivalent of [...]

Exactly, that's what I need open_memmap for, and I often had to search some of my old code for it when I forgot the name again.
@charris: Is this possible with np.load? If so, I'll gladly change my PR to document how!

[charris]

Can't hurt. Thanks @f0k .

[f0k]

Thanks for merging!