TIFFFillTile: Read error when writing a TIFF file on a per-tile basis.

[MartinBuessemeyer]

Hello People,

firstly, thanks a lot for providing the python bindings for vips. They are really helpful 😄.

I am facing a library issue when trying to write a tiff image in tile-based fashion.

The use case:
I get tiles generated by some kind of source (e.g. a CV deep learning model). Since the composed image would be too large to fit in RAM, I want to write it to disk on a per-tile basis, but when reading the image again, I am getting a TIFFFillTile: Read error at row 896, col 896, tile 16; got 0 bytes, expected 49152.

I broke down the issue to a minimal code example:

def test_isolate_error():
    from tempfile import NamedTemporaryFile
    import pyvips as vips

    def numpy_to_vips(array: np.ndarray) -> vips.Image:
        array = array.transpose((1, 0, 2))
        h,w,c = array.shape
        array = array.ravel()
        return vips.Image.new_from_memory(
            data=array.data,
            width=w,
            height=h,
            bands=c,
            format="uchar",
        )
    tiff_size = 1024
    tile_size = 512
    tiling = True
    path = Path(NamedTemporaryFile(suffix=".tiff").name)

    vips_img = vips.Image.black(tiff_size, tiff_size, bands=3)
    vips_img.write_to_file(str(path), tile=tiling)

    for x in range(0, tiff_size, tile_size):
        for y in range(0, tiff_size, tile_size):
            region_img = numpy_to_vips(np.ones((tiff_size, tiff_size, 3), dtype=np.uint8) * 255)
            vips_img.insert(region_img, x, y)
            vips_img.write_to_file(str(path), tile=tiling)
            vips_img = vips.Image.new_from_file(str(path))

This code is failing for me in the latest ubuntu docker image with the most recent vips release with the following error message:

Error Message:
>           raise Error('unable to call {0}'.format(operation_name))
E           pyvips.error.Error: unable to call VipsForeignSaveTiffFile
E             TIFFFillTile: Read error at row 896, col 896, tile 16; got 0 bytes, expected 49152

Of course, it might be the case that my code for writing the image on a per-tile basis is just wrong, and I would need to write it differently. In both cases, help would be appreciated 👍.

Thanks for your time,
Martin

[jcupitt]

Hi @MartinBuessemeyer,

Sorry, I don't think this is going to work :( The best you can do with pyvips is something like (untested!!):

tiles = [some expression to yield a list of numpy tiles]
# you can use `new_from_array` to make a vips image from a numpy array
# they will share memory buffers, so there's no copying and it should be efficient
image = pyvips.Image.arrayjoin([pyvips.Image.new_from_array(tile) for tile in tiles],
                               across=tile_across)
image.write_to_file("x.tif", tile=True, pyramid=True)

But this will unfortunately force the whole image into ram. You could test it and see how memory performance scales, but I fear it may not be good enough.

We've been talking about a better way to generate images like this, but it's not trivial and no one's done the work.

[MartinBuessemeyer]

Thanks for the quick response and your help. I'll give it a try 👍.
Feel free to close the issue if you consider this as done :).

[Himanshunitrr]

@jcupitt any updates on this?

We've been talking about a better way to generate images like this, but it's not trivial and no one's done the work.
Can you suggest some approach regarding how it can be done? More than happy to brainstorm and contribute to this

[jcupitt]

You'd need to add a new operator to libvips which would fire a signal (eg. prepare) to fill a tile. It should probably have an optional mutex to make these signals single-threaded. Something like black would make a good starting point.

You'd then wrap this in pyvips and connect the signal to a method called on_prepare(region) which would be responsible for somehow generating the pixels in that area.

I've no idea what performance would be like. We'd need to do some experiments.