[minor suggestion] optional importing of datasets?

[bluenote10]

import arraymancer was just broken for me due to an issue with my local untar version. I got curious why Arraymancer actually needs this and noticed that the default import builds things like imdb and mnist.

Maybe the included test datasets could be omitted from the default import to make default builds more slim?

[mratsim]

Yes, agreed.

I'm not too sure how I can do CI on optional dependencies though, it would need nimble support nim-lang/nimble#482.

For example HDF5 support is optional but due to this it's not tested :/.

Arraymancer/src/io/io.nim

Lines 5 to 17 in 25cf5e3

	template canImport(x: untyped): bool =
	compiles:
	import x
	import ./io_csv, ./io_npy, ./io_image
	export io_csv, io_npy, io_image
	when canImport(nimhdf5):
	# only provide the hdf5 interface, if the hdf5 library is
	# installed to avoid making `nimhdf5` a dependency
	import ./io_hdf5
	export io_hdf5

[bluenote10]

I'm not too sure how I can do CI on optional dependencies though

I think as a first step keeping them as nimble dependencies is fine. What would already help is to move the import into individual tests instead of letting import arraymancer lead to importing imdb/mnist.

[Vindaar]

Aside from the fact that HDF5 support is not being tested, the canImport solution has its downsides too. Since I obviously have nimhdf5 installed, every time I import arraymancer, I get a nice dependency on HDF5, even though the project itself might not want to touch HDF5 at all.

In my own projects I added an optional compiler flag -d:pure if I want to avoid importing such modules. But that's also not a perfect solution.