qurator-spk/eynollah
1
0
Fork 0
mirror of https://github.com/qurator-spk/eynollah.git synced 2026-05-26 07:39:22 +02:00
Code Issues Projects Releases Wiki Activity
1558 commits 26 branches 23 tags 30 MiB
Commit graph

1558 commits

This branch
This branch
All branches
Author SHA1 Message Date
Robert Sachunsky
bcec0c4a55
Merge 9801129aa6 into 1df32eba87 2026-05-22 10:38:11 +00:00
Robert Sachunsky
9801129aa6 estimate_skew_contours: ensure retval is always float 2026-05-22 12:37:07 +02:00
Robert Sachunsky
26afc5ddab ModelZoo: ensure exported TensorShape is converted to plain tuple 2026-05-22 12:35:44 +02:00
Robert Sachunsky
0836230c6b utils_ocr: avoid module-level import of TF 2026-05-21 22:50:53 +02:00
Robert Sachunsky
f3a93983c0 ModelZoo: add ocr key for memory_limit 2026-05-21 22:50:13 +02:00
Robert Sachunsky
ea41dcae1d trocr: use beam search instead of greedy decoding 2026-05-21 17:52:27 +02:00
Robert Sachunsky
074753a98e ModelZoo: fix Torch device selection 2026-05-21 17:25:53 +02:00
Robert Sachunsky
000e4ac8d8 trocr: extract confidence, too 2026-05-21 17:25:39 +02:00
Robert Sachunsky
f3649adbf2 trocr: apply do_not_mask_with_textline_contour here, too 2026-05-21 17:23:11 +02:00
Robert Sachunsky
1d67e65f11 trocr: simplify, batch over entire page… 
- batching over entire page instead of region-wise
  (underfilling batches)
- avoid copied redundant code
 2026-05-21 15:48:21 +02:00
Robert Sachunsky
d50bd7c650 trocr: avoid warnings by passing clean_up_tokenization_spaces=False 2026-05-21 14:20:51 +02:00
Robert Sachunsky
f9f9130dbb do_order_of_regions: remove redundant+overcautious assertion 2026-05-21 03:21:36 +02:00
Robert Sachunsky
bf7ec0233d ModelZoo.load_model: use memory_limit instead of memory_growth… 
- growth strategy is more flexible, but uses much more VRAM
- limit strategy needs to be calibrated to models (currently fixed),
  and batch size, but needs much less VRAM and is faster
 2026-05-21 02:43:34 +02:00
Robert Sachunsky
94a5e9da14 ModelZoo.load_model: avoid attempting to load exported models as Keras 
models (which causes a warning), but switch to TF-Serving import right away
 2026-05-21 02:41:19 +02:00
Robert Sachunsky
7f2bf715df ModelZoo.load_model: fix loading exported vs saved models 2026-05-21 02:39:59 +02:00
Robert Sachunsky
3de1407d18 drop unnecessary TF / Torch imports 2026-05-21 02:38:20 +02:00
Robert Sachunsky
bdfebd2c70 reload_weights: save() → export() w/ serve() inference 2026-05-19 03:40:18 +02:00
Robert Sachunsky
86adaf299a training.models.transformer_block: tf.reshape → Keras Reshape layer 2026-05-19 03:40:16 +02:00
Robert Sachunsky
9efce5e9f2 Predictor.shutdown: use join() instead of terminate() 2026-05-19 03:40:07 +02:00
Robert Sachunsky
ffe5cdc519 ModelZoo.shutdown: drop extra del (already done by shutdown()) 2026-05-19 03:40:05 +02:00
Robert Sachunsky
481c286da9 ModelZoo.load_model: no XLA compilation 2026-05-19 03:40:05 +02:00
Robert Sachunsky
f329e10a80 test_layout: rm ignored --allow_scaling option 2026-05-19 03:40:04 +02:00
Robert Sachunsky
17b311441a model_zoo: also parse comma/colon syntax for device in Torch case 2026-05-19 03:40:03 +02:00
Robert Sachunsky
be4fe8c263 contour: drop unused functions depending on rotation_image_new() 2026-05-19 03:40:02 +02:00
Robert Sachunsky
87cce6c963 CLI tests: add opt-in envvar EYNOLLAH_OPTIONS for device selection, 
model directory etc.
 2026-05-19 03:40:01 +02:00
Robert Sachunsky
1ed633bc25 test_model_zoo: adapt (load_models instead of load_model) 2026-05-19 03:40:00 +02:00
Robert Sachunsky
21ecb043f7 CLIs: move --device option to group level 2026-05-19 03:39:59 +02:00
Robert Sachunsky
7ed1a1ebac CLIs: allow -h and show defaults uniformly, harmonise help, drop 
remaining redundant negative options
 2026-05-19 03:39:56 +02:00
Robert Sachunsky
cd62f13872 eynollah_ocr: make work again, re-use Eynollah base class… 
- re-use Eynollah base class
- use `ModelZoo.load_models()` instead of `load_model()`
- pass in `device` init kwarg, delegate to `ModelZoo.load_models()`
- `device`: return Torch device at loaded model tensors
  instead of ad-hoc selection
- make numeric init kwargs non-optional (only numeric)
 2026-05-19 03:39:55 +02:00
Robert Sachunsky
ded668a256 model_zoo: fix clash between Predictor and direct (OCR) use-cases… 
- `load_models()`: uniformly handle arg types
- `load_model()`: move handling of non-model categories
  to `load_models()`
- `load_model()`: move SavedModel preference over HDF5 to `model_path()`
- `_load_ocr_model()`: add user-selected device handling and reporting
  for Torch (as for TF)
- `_load_ocr_model()`: move (TF-based) CNN-RNN case to `load_model()`
  (including Keras layer mapping)
- `shutdown()`: only apply `shutdown()` to Predictor model types
 2026-05-19 03:39:53 +02:00
Robert Sachunsky
98e6fbbcbb mbreorder: make work again, re-use Eynollah base class 2026-05-19 03:39:52 +02:00
Robert Sachunsky
7e8b9311d3 Revert "test_model_zoo: fix calls" 
This reverts commit 5a98f55be3.
 2026-05-19 03:32:37 +02:00
Robert Sachunsky
a1449da1d1 Revert "fix model loading in mb_ro and ocr" 
This reverts commit 218a95e6a0.
 2026-05-19 03:32:19 +02:00
kba
1df32eba87 CD: base docker image: typo {,v}3.13.0 2026-05-11 13:41:30 +02:00
kba
d7337a3080 CD: base docker image on versioned ocrd/core-cuda-tf2:v3.13.0 2026-05-11 13:38:36 +02:00
kba
e612db2bb1 📦 v0.8.0 2026-05-11 13:16:30 +02:00
kba
6cfbd93ac7 📝 changelog 2026-05-11 13:14:56 +02:00
kba
c7104c2852 Merge branch 'prepare-release-v0.8.0' 2026-05-11 13:12:19 +02:00
kba
5a98f55be3 test_model_zoo: fix calls 2026-05-11 12:22:24 +02:00
kba
218a95e6a0 fix model loading in mb_ro and ocr 2026-05-11 12:19:20 +02:00
kba
2035b07b55 Merge remote-tracking branch 'bertsky/ro-fixes-final' into prepare-release-v0.8.0 
# Conflicts:
#	requirements-ocr.txt
 2026-05-11 09:46:17 +02:00
Robert Sachunsky
db87aa995d reqs for OCR: relax ad5f2272 (depending on Python version) 2026-05-11 03:15:54 +02:00
Robert Sachunsky
e183937c5d separate_lines_new2: fix coord overflow by clipping, simplify… 
- found positive and negative peaks, and even more so their
  relative offsets, may overflow in the cropped image,
  causing fake textlines; avoid that by clipping to the valid
  y coordinates
- calculation for number of tiles: sometimes one less
  tile is needed by making the previous last tile
  half-full on the right side
- add some (commented) plotting
- simplify (a lot, but only partially)
 2026-05-11 03:09:02 +02:00
Robert Sachunsky
130f0aee42 do_work_of_slopes_curved: improve on d257869d… 
- relative images now need larger relative min_area
  (i.e. compensation factors)
- do not attempt (even) single-line skew estimation
  (via linear regression) if there is no (large enough)
  contour at all
- avoid re-computing `mask_parent`
- add some (commented) plotting
 2026-05-11 03:03:04 +02:00
kba
ce5d6bc43c try to accomodate outdaten Python versions unsupported by current transformers 2026-05-09 18:03:40 +02:00
kba
03f3f9af17 update model zoo and docs to link to v0_8_0 model release on zenodo 2026-05-09 17:58:59 +02:00
Robert Sachunsky
a61fb09ec5 CI: drop py3.8 (u/a for new req transformers >= 5) 2026-05-09 04:14:49 +02:00
Robert Sachunsky
4406a0299e update CLI test for binarization… 
- update expected log messages
 2026-05-09 04:12:19 +02:00
Robert Sachunsky
4cd398bd0d standalone binarization: update, simplify… 
- re-use Eynollah base class, drop copied code
- simplify `run()` and `run_single()`
- delegate to `do_prediction()`
  instead of custom (old) tiling loop
- drop `predict()`
- add `--device` option to CLI as well
 2026-05-09 04:12:02 +02:00
Robert Sachunsky
29abae0144 update CLI test for enhancer… 
- update expected log messages
- force `-ncu 3`, because otherwise
  the example images would not be deemed
  in need of enhancement
 2026-05-09 02:59:52 +02:00