Missing Models

[Schneipi]

Since there are no models hosted on https://ocr-d-repo.scc.kit.edu, does anyone know where we can access them now? More specifically, I would want to use:

• latest_net_G.pth
• block_segmentation_weights.h5
• mapping_densenet.pickle

The URLs are outdated here: https://github.com/OCR-D/core/blob/master/ocrd/ocrd/resource_list.yml

[cneud]

@Schneipi I've added the trained models to the latest release here:
https://github.com/OCR-D/ocrd_anybaseocr/releases/tag/v1.8.2.

[Schneipi]

@cneud thanks!

What about mapping_densenet.pickle? It seems to be required for ocrd-anybaseocr-layout-analysis. Running ocrd resmgr download ocrd-anybaseocr-layout-analysis '*' does not download the file and therefore ocrd-anybaseocr-layout-analysis throws an UnpicklingError.

seg_model.hdf5, provided with the latest release, seems to concern ocrd-anybaseocr-tiseg, if I'm not mistaken.

[cneud]

@Schneipi You are right, there seem to be more required models ;)

For mapping_densenet.pickle, there is a Tensorflow SavedModel version for this according to resource_list.yml at https://ocr-d.kba.cloud/structure_analysis.tar.gz (but which I am not sure is identical/working).

[Edit: as it turns out, structure_analysis.h5 is just another required model, but no replacement for mapping_densenet.pickle.]

Unfortunately I don't have a copy of mapping_densenet.pickle from https://ocr-d-repo.scc.kit.edu/models/dfki/layoutAnalysis/mapping_densenet.pickle...

[kba]

We do have a copy of all the models, @tdoan2010 knows where to find them, we will find a more permanent place for them and update the resource_list.yml later this week, I'll keep you posted.

[tdoan2010]

Hi, yes, I have a copy of all models. And I guess here is the mapping_densenet.pickle that you're looking for.

[cneud]

That mapping_densenet.pickle looks suspicious as the file size is less than 1 kB...

[kba]

That mapping_densenet.pickle looks suspicious as the file size is less than 1 kB...

IIRC this was just pickled dict mapping network outputs to region types, so the small file size tracks.

[cneud]

IIRC this was just pickled dict mapping network outputs to region types

I see, thx. Then I will also add the file to the release.