Dimension mismatch with loading a lazily loaded image to memory

[RpfR2000]

I have loaded in a large image (26 GB) lazily with the lazyio=true flag and am now trying to load a subset of that image to memory by indexing as follows:

chunk = Array{Gray{N0f16}, 3}(undef, y_end-y_start+1, x_end-x_start+1, slices)
for k in 1:slices
    chunk[:,:,k] .= image[y_start:y_end, x_start:x_end, k]
end

This gives the following error: LoadError: DimensionMismatch: parent has 18596 elements, which is incompatible with size (256, 256). The reason I am loading in a loop like this is because I couldn't even @show the size of a 3D chunk of the image without the same error being thrown. I can @show the size of a slice, though (not sure why). I have verified that the size and eltype of the lazily loaded image is correct. Any ideas what is going on?

[tlnagy]

Does image[y_start:y_end, x_start:x_end, k] give you the dimensions and eltype of what you expect? Have you tried loading the image with mmap=true instead of lazyio?

[RpfR2000]

Yes, image[y_start:y_end, x_start:x_end, k] gives the correct eltype and dimensions. I just tried mmap=true and got LoadError: BoundsError: attempt to access 0-element Vector{TiffImages.Tag} at index [1].

[tlnagy]

Interesting. Can you give me the output of first(ifds(img))?

[RpfR2000]

first(ifds(image)) = IFD, with tags:
Tag(SUBFILETYPE, 2)
Tag(IMAGEWIDTH, 10138)
Tag(IMAGELENGTH, 10138)
Tag(BITSPERSAMPLE, 16)
Tag(COMPRESSION, COMPRESSION_NONE)
Tag(PHOTOMETRIC, 1)
Tag(IMAGEDESCRIPTION, "<?xml version="1.0" ...")
Tag(SAMPLESPERPIXEL, 1)
Tag(SOFTWARE, "OME Files (C++) 0.6....")
Tag(TILEWIDTH, 256)
Tag(TILELENGTH, 256)
Tag(TILEOFFSETS, UInt64[16, 131088, 262160, 393232, 524304, ...])
Tag(TILEBYTECOUNTS, UInt64[131072, 131072, 131072, 131072, 131072, ...])
Tag(SAMPLEFORMAT, 1)
Tag(UNKNOWN(65325), 1329.7020542975515)
Tag(UNKNOWN(65326), 0.065)
Tag(UNKNOWN(65327), 1.0)
Tag(UNKNOWN(65328), 1.0)
Tag(UNKNOWN(65329), Float64[-13174.5, -14112.900000000001, 4663.08, 1329.7020542975515])

[tlnagy]

Ah, the image is tiled. mmap doesn't support tiled images if I remember properly, but lazyio should, so let's see if we can diagnose this.

This gives the following error: LoadError: DimensionMismatch: parent has 18596 elements, which is incompatible with size (256, 256).

This error now makes more sense, since the tile sizes are 256x256. Do you mind providing the full stacktrace (all of the output) for lazyio=true error? I'm curious where it's failing.

[tlnagy]

I suspect the error is being triggered because it's unaware that the image is tiled. Actually it looks like it should set the cache size properly

TiffImages.jl/src/layout.jl

Lines 138 to 141 in f4d285b

	if istiled(ifd)
	tile_width = tilecols(ifd)
	tile_height = tilerows(ifd)
	return Array{pixeltype}(undef, cld(ncols(ifd), tile_width) * tile_width, cld(nrows(ifd), tile_height) * tile_height)

If you lazyio load the image, what's the output of size(img.cache) and size(img)

[RpfR2000]

size(img.cache) doesn't work, I get type DenseTaggedImage has no field cache.

size(image) = (10138, 10138, 135)

Full stacktrace:

ERROR: LoadError: DimensionMismatch: parent has 18596 elements, which is incompatible with size (256, 256)
Stacktrace:
  [1] _throw_dmrs(n::Int64, str::String, dims::Tuple{Int64, Int64})
    @ Base ./reshapedarray.jl:182
  [2] _reshape
    @ ./reshapedarray.jl:177 [inlined]
  [3] reshape
    @ ./reshapedarray.jl:112 [inlined]
  [4] reshape
    @ ./reshapedarray.jl:117 [inlined]
  [5] _broadcast_getindex_evalf
    @ ./broadcast.jl:709 [inlined]
  [6] _broadcast_getindex
    @ ./broadcast.jl:682 [inlined]
  [7] _getindex
    @ ./broadcast.jl:706 [inlined]
  [8] _broadcast_getindex
    @ ./broadcast.jl:681 [inlined]
  [9] getindex
    @ ./broadcast.jl:636 [inlined]
 [10] macro expansion
    @ ./broadcast.jl:1004 [inlined]
 [11] macro expansion
    @ ./simdloop.jl:77 [inlined]
 [12] copyto!
    @ ./broadcast.jl:1003 [inlined]
 [13] copyto!
    @ ./broadcast.jl:956 [inlined]
 [14] copy
    @ ./broadcast.jl:928 [inlined]
 [15] materialize
    @ ./broadcast.jl:903 [inlined]
 [16] tile(cache::Matrix{Gray{N0f16}}, ifd::TiffImages.IFD{UInt64})
    @ TiffImages ~/.julia/packages/TiffImages/Xrmi4/src/load.jl:151
 [17] transform
    @ ~/.julia/packages/TiffImages/Xrmi4/src/load.jl:139 [inlined]
 [18] getindex(A::LazyBufferedTIFF{Gray{N0f16}, UInt64, Matrix{Gray{N0f16}}}, i1::Int64, i2::Int64, i::Int64)
    @ TiffImages ~/.julia/packages/TiffImages/Xrmi4/src/types/lazy.jl:129
 [19] macro expansion
    @ ./multidimensional.jl:917 [inlined]
 [20] macro expansion
    @ ./cartesian.jl:64 [inlined]
 [21] macro expansion
    @ ./multidimensional.jl:912 [inlined]
 [22] _unsafe_getindex!
    @ ./multidimensional.jl:925 [inlined]
 [23] _unsafe_getindex(::IndexCartesian, ::LazyBufferedTIFF{Gray{N0f16}, UInt64, Matrix{Gray{N0f16}}}, ::UnitRange{Int64}, ::UnitRange{Int64}, ::Int64)
    @ Base ./multidimensional.jl:903
 [24] _getindex
    @ ./multidimensional.jl:889 [inlined]
 [25] getindex
    @ ./abstractarray.jl:1291 [inlined]
 [26] getindex
    @ ~/.julia/packages/TiffImages/Xrmi4/src/types/dense.jl:67 [inlined]

[tlnagy]

size(img.cache) doesn't work, I get type DenseTaggedImage has no field cache.

That means that you weren't loading using with lazyio=true, do you mind trying with that flag?

[RpfR2000]

This was the code I ran:

image = TiffImages.load(path*filename; lazyio=true) 
@show size(image.cache)
@show size(image)

[tlnagy]

Ah, my bad. I forgot that we wrap with a DenseTaggedImage, could you give me the output of size(img.data.cache)?

[RpfR2000]

size(image.data.cache) = (10240, 10240)

[RpfR2000]

@tlnagy perhaps I could also share the image with you if that would be easier? It is quite large though.

[tlnagy]

Yeah, do you mind uploading it to dropbox/similar place and I can download it? It'll be easier to diagnose that way then going back and forth here.

You can send the link to my email listed on my github profile page.

[RpfR2000]

Sent

[tlnagy]

Do you mind sharing it as a public link? It's requiring me to log in to access it.

[RpfR2000]

Sorry, just sent a link. Let me know if that works.

[chrstphrbrns]

I'm preparing a PR for this

[RpfR2000]

Does this require a PR?

[tlnagy]

Thanks to @chrstphrbrns' quick work this should be fixed in the newest release v0.11.2, give it a hour or so for the version to propagate through CI etc.