Commit graph

  • a9523d78f4
    Merge 12b1271487 into c9f6aa35b2 Robert Sachunsky 2026-04-12 23:20:20 +00:00
  • 12b1271487 layout cli: add option --halt-fail Robert Sachunsky 2026-04-13 01:19:47 +02:00
  • 56e6deb02c predictor: jit-compile and precompile (non-autosized) models Robert Sachunsky 2026-04-13 01:17:04 +02:00
  • 01c54eb2ef reduce inference batch sizes to accommodate 8 GB VRAM Robert Sachunsky 2026-04-13 01:15:25 +02:00
  • f44c39667e predictor: disable rebatching (until we have flexible batch sizes) Robert Sachunsky 2026-04-13 01:14:49 +02:00
  • 219954d15b predictor: use predict_on_batch instead of predict Robert Sachunsky 2026-04-13 01:14:18 +02:00
  • 0d21b62aee disable autosized prediction entirely (also for _patched)… Robert Sachunsky 2026-04-10 18:23:10 +02:00
  • ccef63f08b get_regions: always use resized/enhanced image… Robert Sachunsky 2026-04-10 18:17:51 +02:00
  • 04da66ed73 training: plot only ~ 1000 training and ~ 100 validation images Robert Sachunsky 2026-03-30 13:34:05 +02:00
  • a8556f5210 run: sort parallel log messages by file name instead of prefixing… Robert Sachunsky 2026-03-30 13:18:40 +02:00
  • 62140e4159
    Merge 9858221724 into c9f6aa35b2 Konstantin Baierer 2026-03-27 08:37:01 +00:00
  • 9858221724 comment out printing file names while training cnn-rnn ocr model integrating_trocr_and_torch_ensembling_and_updating_characters_list vahidrezanezhad 2026-03-27 09:36:55 +01:00
  • 8333158ecc BUG fixing for cnn-rnn ocr model training if augmentation is false vahidrezanezhad 2026-03-27 09:15:19 +01:00
  • 1756443605 fixup device sel Robert Sachunsky 2026-03-16 15:35:07 +01:00
  • bd495279e2
    Merge 42a3cc2335 into c9f6aa35b2 Robert Sachunsky 2026-03-16 11:32:07 +00:00
  • 6bbdcc39ef CLI/Eynollah.setup_models/ModelZoo.load_models: add device option/kwarg Robert Sachunsky 2026-03-15 04:54:04 +01:00
  • 67e9f84b54 do_prediction* for "col_classifier": pass array as float16 instead of float64 Robert Sachunsky 2026-03-15 03:20:39 +01:00
  • f54deff452 model_zoo/predictor: use one subprocess per model… Robert Sachunsky 2026-03-15 02:53:37 +01:00
  • c514bbc661 make switching between autosized and looped tiling easier Robert Sachunsky 2026-03-14 02:16:26 +01:00
  • 2f3b622cf5 predictor: rebatch tasks to increase CUDA throughput… Robert Sachunsky 2026-03-14 00:52:34 +01:00
  • b550725cc5 wrap_layout_model_patched: simplify shape calculation Robert Sachunsky 2026-03-14 00:51:22 +01:00
  • d6404dbbc2 do_prediction*: pass arrays as float16 instead of float64 to TF Robert Sachunsky 2026-03-14 00:49:26 +01:00
  • 135064a48e model_zoo: region model not used at runtime anymore - don't load Robert Sachunsky 2026-03-14 00:48:52 +01:00
  • ec08004fb0 run: add QueueListener to pool / QueueHandler to workers… Robert Sachunsky 2026-03-14 00:43:58 +01:00
  • b7aa1d24cc CLI: drop redundant negative option forms, add --num-jobs Robert Sachunsky 2026-03-13 18:22:25 +01:00
  • 576e120ba6 autosized prediction is only faster for _patched, not for _resized… Robert Sachunsky 2026-03-13 18:15:30 +01:00
  • 7499e3e7b8 textline inference thresholding was disabled during the merging step vahidrezanezhad 2026-03-13 17:48:27 +01:00
  • 6d55f297a5 run: use ProcessPoolExecutor for parallel run_single across pages… Robert Sachunsky 2026-03-13 10:15:51 +01:00
  • 96cfddf92d split_textregion_main_vs_header: avoid zero division Robert Sachunsky 2026-03-13 02:41:06 +01:00
  • 4e9b062b84 separate_marginals_to_left_and_right...: simplify Robert Sachunsky 2026-03-13 02:40:33 +01:00
  • ae0f194241 drop ProcessPoolExecutor for intra-page parallel subprocessing… Robert Sachunsky 2026-03-13 02:38:40 +01:00
  • becf031c65 refactor to remove data-dependency from all Eynollah methods… Robert Sachunsky 2026-03-13 01:44:39 +01:00
  • 800c55b826 predictor: fix spawn vs fork / parent vs child contexts Robert Sachunsky 2026-03-13 02:42:16 +01:00
  • 64281768a9 run_graphics_and_columns_light: fix double 1-off error… Robert Sachunsky 2026-03-12 10:18:14 +01:00
  • 46c5f52491 CLI: don't append /models_eynollah here (already in default_specs) Robert Sachunsky 2026-03-11 02:39:32 +01:00
  • 10214dfdda predictor: make sure all shared arrays get freed eventually Robert Sachunsky 2026-03-11 02:38:11 +01:00
  • cf5caa1eca predictor: fix termination for pytests… Robert Sachunsky 2026-03-11 02:34:29 +01:00
  • bb468bf68f predictor: mp.Value must come from spawn context, too Robert Sachunsky 2026-03-11 02:27:47 +01:00
  • 9f127a0783 introduce predictor subprocess for exclusive GPU processing… Robert Sachunsky 2026-03-07 03:54:16 +01:00
  • 6f4ec53f7e wrap_layout_model_resized/patched: compile call instead of predict Robert Sachunsky 2026-03-07 03:52:14 +01:00
  • 338c4a0edf wrap layout models for prediction (image resize or tiling) all in TF Robert Sachunsky 2026-03-07 03:33:44 +01:00
  • f33fd57da8 model_zoo: resolve path names coming in from caller (CLI) Robert Sachunsky 2026-03-05 00:45:24 +01:00
  • 41dccb216c use (generalized) do_prediction() instead of predict_enhancement() Robert Sachunsky 2026-03-04 23:49:11 +01:00
  • 341480e9a0 do_prediction: if img was too small for model, also upscale results Robert Sachunsky 2026-03-04 23:41:45 +01:00
  • 8ebbe65c17 textline_contours: remove unnecessary resize_image, simplify Robert Sachunsky 2026-03-04 15:13:34 +01:00
  • 3370a3aa85 do_prediction*: avoid 3-channel results, simplify further… Robert Sachunsky 2026-03-03 01:20:16 +01:00
  • f1d8257496 page alto label generation activated for textline vahidrezanezhad 2026-03-03 21:12:20 +01:00
  • 4b80e45d91 character list only needs be copied for cnn-rnn ocr model vahidrezanezhad 2026-03-03 13:20:22 +01:00
  • c9f6aa35b2
    fix license badge main Clemens Neudecker 2026-03-03 09:43:54 +01:00
  • ff7dc31a68 do_prediction*: rename identifiers for artificial class thresholding Robert Sachunsky 2026-03-02 13:08:11 +01:00
  • b9cf68b51a training: fix b6d2440c Robert Sachunsky 2026-03-01 20:00:05 +01:00
  • ae3b6916ee assert within vit_resnet50_unet model is commented out since arising assert error vahidrezanezhad 2026-03-01 18:39:30 +01:00
  • 7f7bdab208 patches class for VIT encoder is corrected vahidrezanezhad 2026-03-01 18:26:29 +01:00
  • 686f1d34aa do_prediction*: simplify (esp. indexing/slicing) Robert Sachunsky 2026-03-01 04:37:20 +01:00
  • 3b56fa2a5b training: plot GT/prediction and metrics before training (commented) Robert Sachunsky 2026-02-28 20:08:10 +01:00
  • e47653f684 training: move nCC metric/loss to .metrics and rename… Robert Sachunsky 2026-02-28 20:01:49 +01:00
  • 361d40c064 training: improve nCC metric/loss - measure localized congruence… Robert Sachunsky 2026-02-28 19:44:10 +01:00
  • 7e06ab2c8c training: add config param add_ncc_loss for layout/binarization… Robert Sachunsky 2026-02-27 12:55:15 +01:00
  • c6d9dd7945 training: use mixed precision and XLA (commented; does not work, yet) Robert Sachunsky 2026-02-27 12:57:47 +01:00
  • c1d8a72edc training: shuffle tf.data pipelines Robert Sachunsky 2026-02-28 20:04:32 +01:00
  • 1cff937e72 training: make data pipeline in 7888fa5 more efficient Robert Sachunsky 2026-02-27 12:53:09 +01:00
  • f8dd5a328c training: make plotting 18607e0f more efficient… Robert Sachunsky 2026-02-27 12:50:37 +01:00
  • fed005abd7 Fix a bug by saving model weights in steps vahidrezanezhad 2026-02-28 09:01:06 +01:00
  • 2d5de8e595 training.models: use bilinear instead of nearest upsampling… Robert Sachunsky 2026-02-27 12:48:28 +01:00
  • ba954d6314 training.models: fix daa084c3 Robert Sachunsky 2026-02-27 12:47:59 +01:00
  • 7c3aeda65e training.models: fix 9b66867c Robert Sachunsky 2026-02-27 12:40:56 +01:00
  • 439ca350dd training: add metric ConfusionMatrix and plot it to TensorBoard Robert Sachunsky 2026-02-26 13:55:37 +01:00
  • b6d2440ce1 training.utils.preprocess_imgs: fix polymorphy in 27f43c1 Robert Sachunsky 2026-02-25 20:39:15 +01:00
  • 42bab0f935 docs/train: document --missing-printspace=project Robert Sachunsky 2026-02-25 13:18:40 +01:00
  • 4202a1b2db training.generate-gt.pagexml2label: add --missing-printspace Robert Sachunsky 2026-02-25 11:16:21 +01:00
  • 7823ea2c95 training.train: add early stopping for OCR Robert Sachunsky 2026-02-25 00:16:07 +01:00
  • 36e370aa45 training.train: add validation data for OCR Robert Sachunsky 2026-02-25 00:10:43 +01:00
  • b399db3c00 training.models: simplify CTC loss layer Robert Sachunsky 2026-02-24 20:43:50 +01:00
  • 92fc2bd815 training.train: fix data batching for OCR in 27f43c17 Robert Sachunsky 2026-02-24 20:42:08 +01:00
  • 86b009bc31 training.utils.preprocess_imgs: fix file name stemming 27f43c17 Robert Sachunsky 2026-02-24 20:41:08 +01:00
  • 20a3672be3 training.utils.preprocess_imgs: fix file shuffling in 27f43c17 Robert Sachunsky 2026-02-24 20:37:44 +01:00
  • 658dade0d4 training.config_params: flip_index needed for scaling_flip, too Robert Sachunsky 2026-02-24 20:36:00 +01:00
  • abf111de76 training: add metric for (same) number of connected components Robert Sachunsky 2026-02-24 17:03:21 +01:00
  • 18607e0f48 training: plot predictions to TB logs along with training/testing Robert Sachunsky 2026-02-24 17:00:48 +01:00
  • 56833b3f55 training: fix data representation in 7888fa5 Robert Sachunsky 2026-02-24 16:46:19 +01:00
  • d8667f46d7 musicregion is added to pagexml to label vahidrezanezhad 2026-02-24 15:46:15 +01:00
  • 76ac4c5b71 Amiri font which works for both arabic and latin vahidrezanezhad 2026-02-24 13:55:45 +01:00
  • 0ca2d02ee8 get label for decoration without type attribute vahidrezanezhad 2026-02-24 01:39:12 +01:00
  • c4434c7f7d same task name for transformer-ocr training and inference vahidrezanezhad 2026-02-19 13:59:16 +01:00
  • a84ae67e7a fix a typo vahidrezanezhad 2026-02-19 00:04:42 +01:00
  • 77adcbea8a copy characters list needed for cnn-rnn ocr model output while training and ensembling vahidrezanezhad 2026-02-18 16:47:21 +01:00
  • 4f66734e4d eynollah config files has renamed from config.json to config_eynollah.json - training trocr model still misses to write config file into checkpoint directories vahidrezanezhad 2026-02-18 16:04:44 +01:00
  • b426f7f152 trocr inference is integrated - works on CPU cause seg fault on GPU vahidrezanezhad 2026-02-18 15:04:54 +01:00
  • a9496bbc70 enhancer/mbreorder: use std Keras data loader for classification Robert Sachunsky 2026-02-17 18:39:30 +01:00
  • 003c88f18a fix double import in 82266f82 Robert Sachunsky 2026-02-17 18:23:32 +01:00
  • f61effe8ce fix typo in c8240905 Robert Sachunsky 2026-02-17 18:20:58 +01:00
  • 5f71333649 fix missing import in 49261fa9 Robert Sachunsky 2026-02-17 18:11:49 +01:00
  • 67fca82f38 fix missing import in 27f43c17 Robert Sachunsky 2026-02-17 18:06:08 +01:00
  • 6a4163ae56 fix typo in 27f43c17 Robert Sachunsky 2026-02-17 17:48:43 +01:00
  • c1b5cc92af fix typo in 7562317d Robert Sachunsky 2026-02-17 17:43:57 +01:00
  • 7bef8fa95a training.train: add verbose=1 consistently Robert Sachunsky 2026-02-17 17:24:07 +01:00
  • 9b66867c21 training.models: re-use transformer builder code Robert Sachunsky 2026-02-17 17:35:20 +01:00
  • daa084c367 training.models: re-use UNet decoder builder code Robert Sachunsky 2026-02-17 17:11:15 +01:00
  • fcd10c3956 training.models: re-use RESNET50 builder (+weight init) code Robert Sachunsky 2026-02-17 14:52:04 +01:00
  • 4414f7b89b training.models.vit_resnet50_unet: re-use IMAGE_ORDERING Robert Sachunsky 2026-02-17 14:18:32 +01:00