port processor to core v3

[kba]

With this PR, eynollah supports OCR-D/core#1240. It simplifies it a lot too.

I'll update the ocrd-tool.json with the changed/added flags here as well.

Draft, please don't merge until v3 stable is released

[bertsky]

Thanks – LGTM!

Have not tested yet, though.

Current main also looks very promising – will give it a try myself

[bertsky]

Note: that filename might not be where that image came from in workspace.image_from_page. It could well be a derived image generated by some previous processor (just not a cropped, deskewed or binarized image, because that would have changed its coordinate system).

[kba]

It's still a bit hazy for me when image_filename is actually used. Ideally, image_pil should take preference and image_filename is only for the plotter/writer, at least in the "single image mode" we're using.

One of the aspects I hope I'll be able to improve a bit with https:/qurator-spk/eynollah/tree/refactoring-2024-08/

[bertsky]

Perhaps we can also re-use session across Eynollah invokations in addition to models?

[kba]

In theory, yes, but with standalone eynollah being focused on batch processing now, I am honestly not sure how/where sessions are defined for the non-dir_in option - @vahidrezanezhad can you tell us?

[bertsky]

Suggested change

	# if not('://' in page.imageFilename):
	# image_filename = next(self.workspace.mets.find_files(local_filename=page.imageFilename)).local_filename
	# else:
	# # could be a URL with file:// or truly remote
	# image_filename = self.workspace.download_file(next(self.workspace.mets.find_files(url=page.imageFilename))).local_filename

This whole effort was to ensure we can pass a working local filename, as (was) needed by Eynollah. The approach by OCR-D is Workspace.image_from_page / Workspace.image_from_segment which will search for the right original or derived image, download it if necessary and load it into memory.

I don't recall what the new behaviour of Eynollah is. If both an image filename and an image object are passed, who wins?

Assuming it's the memory object: this can be removed. (But then I wonder why we still pass the image filename at all...)

[kba]

Well, currently we have

            if image_pil:
                self._imgs = self._cache_images(image_pil=image_pil)
            else:
                self._imgs = self._cache_images(image_filename=image_filename)
[...]
     def _cache_images(self, image_filename=None, image_pil=None):
         ret = {}
         if image_filename:
             ret['img'] = cv2.imread(image_filename)
             self.dpi = check_dpi(image_filename)
         else:
             ret['img'] = pil2cv(image_pil)
             self.dpi = check_dpi(image_pil)

image_filename is (should) then only used passively, to generate filenames of plotted debug images as well as for PAGE serialization.

So I think image_pil should win but for now we need both. But as I said above, one of those things I would love to untangle in the refactoring.

[bertsky]

BTW, I just tested under (METS Server and) OCRD_MAX_PARALLEL_PAGES=2 – it works, but you need lots of GPU memory, otherwise GPU OOM happens. (It does work with CUDA_VISIBLE_DEVICES=, but of course the CPU utilization grows, so that might stall the system.)

I'm not sure if this warrants adding max_workers = 1 to EynollahProcessor ...