drop julia 0.5, view keyword arg, and 16-bit tiffs

[bjarthur]

closes #102

have load return a view instead of using permutedims. note that there are further optimizations to be had here, by writing in-place versions of flipdim.

and separately fix saving of 16-bit tiffs. interestingly, even on master the "binary png" and "alpha" tests are failing. have not taken the time to git bisect when this breakage occurred.

[timholy]

I suppose we should do this. The one reason I hesitate here is that I think ImageMagick doesn't do anything special to load truly large files. For example, in NRRD and ImagineFormat we use mmap to load files that can be much, much larger than the available memory. If using ImageMagick, it seems that IM has to be able to load the whole file into a buffer in RAM. As a consequence, the best case scenario for this PR is that it lets you load a 2x larger image. Is that narrow "window" worth it? I worry that even if we make this operation work, it's not unlikely that the very next operation the user executes on the image might fail. In contrast, shopping for more RAM is an effective solution for many problems.

[timholy]

Again, I'm not strongly against this (indeed I might on balance be slightly in favor), but I do have some reluctance to slow down operations for all the 2d computer vision people who work with small images.

[bjarthur]

is it too much to ask those with small files to collect(load(...)) ?

alternatively, what about using a keyword argument to specify a view, like load(filename, view=true), with the default being false.

[timholy]

I like the keyword argument idea, I think that makes sense.

[bjarthur]

keyword arg added.

but oi, PermutedDimsArray is not in Compat.jl it seems, so this PR will not work on julia 0.5. i might just be too lazy to merge until after we drop support.

[bjarthur]

hah, just noticed Images.jl has dropped support for 0.5. should we do the same for ImageMagick?

also, has it been discussed at all to add a FlippedDimArray to julia Base to match flipdim, much like PermutedDimArray matches permutedims? that would nicely round out this PR.

[timholy]

Sure, dropping support for 0.5 seems very reasonable. I bumped all the JuliaImages packages so that I could run femtocleaner to eliminate depwarns on 0.7 (not that I'm using it in practice yet).

With regards to FlippedDimArray, you can get that with things along the lines of view(img, reverse.(indices(img))...).

[bjarthur]

doh! thanks.

flipdim has been replaced with a view (which can be collected with the same view keyword argument).

and julia 0.5 support has been dropped. should remember to bump the minor version number for the next tag

[timholy]

Maybe better to say "set view=true to reduce memory consumption when loading large files, possibly at some slight cost in terms of performance of future operations."

[bjarthur]

done. is there any way to add a doc string for sth like this? i tried adding one in src/ImageMagick.jl, but it didn't work after using FileIO.

[timholy]

Interesting question, since FileIO is not meant to be extended. Maybe put it on ImageMagick.load? (In other words, does ?ImageMagick.load work as a way of adding a docstring for the load function in this specific package, without "contaminating" ?load?)

[bjarthur]

yes, putting in on ImageMagick.load makes ?ImageMagick.load work. but given that one (or something) must do a using ImageMagick and that docstring is not displayed with ?load, is it worth it?

[timholy]

Certainly not necessary, but in the absence of a better idea I'm not sure where to aggregate the IM-specific extensions. We also have fps and perhaps other kwargs.

[timholy]

Not that you should feel responsible for doing this, simply asking the question is useful.

[timholy]

Thanks again!

[bjarthur]

i was planning to add a docstring this morning, but guess will leave it for another time. thanks!