- move `--device` option to group level, apply to all model types (including Torch/ONNX)
- load models w/ `memory_limit` instead of `memory_growth` strategy (faster and less VRAM)
- show full stacktrace in case predictor fails (not just exception name)
- :fire: default to ONNX inference w/ TensorRT instead of TF (much **faster**, but requires **warmup** phase w/ persistent cache directory `$XDG_CONFIG_HOME`)
- :fire: published **new** set of **models**, both for training (TF/Keras) and inference (ONNX) – rebuilt from code changes (see below), *not* retrained
- :fire: Docker image now based on `ocrd/core-cuda-onnx` for layout only, no `[OCR]` in Docker ATM
- OCR: run pages in **parallel** (as for layout) via forking, add `--halt-fail` and `--num-jobs`, too
- improve layout:
* for heuristic reading order, do not try to elongate horizontal separators
* when column classifier is confident enough, do maximally enlarge the image (for 6 columns or more)
- improve TrOCR:
* refactor, simplify
* batch over entire page (faster)
* extract confidence, too
* use beam search instead of greedy decoder
* load model and preprocessor/tokenizer into one object (no need for distinct models)
* no need to resize images in advance
* if set, apply `-nmtc` here, too
* skip lines lower than `-min_conf` instead of setting empty string
- improve Keras OCR:
* refactor, simplify
* batch over entire page (faster)
* use correct confidence estimation
* run binarization ad-hoc (if not provided)
* adapt to all-in-one inference model
* get image size from model (instead of fixed)
* apply `-min_conf` here, too
* skip lines lower than `-min_conf` instead of setting empty string
* separate off `.png` files if `dir_in_bin==dir_in`
* batch flipped line candidates together with normal lines
- training/setup:
* refactor imports from `.models` (single auto-configured `get_model()` call, no `custom_objects` loading)
* drop new setting `reload_weights` in favour of `--rebuild` option for new CLI `eynollah-training convert`
* new CLI for model conversion between Keras (formats HDF5 / native Keras, TF SavedModel), TF-Serving (i.e. `model.export()`) and ONNX
* extract `MusicRegion` from PAGE GT, too
- training/models:
* ViT models: use Keras `Reshape` layer instead of ad-hoc `tf.reshape`
* ViT models: use `tf.map_fn` to iterate over batch in `tf.image.extract_patches` for attention (faster, less VRAM, makes ONNX conversion work)
* Keras (CNN-RNN) OCR backend: replace `Conv1D(channels_first)` (not fully supported by TF/CUDNN on CPU) by `Conv1D(channels_last)` w/ `Permute` layers
* Keras (CNN-RNN) OCR training→inference conversion: encapsulate CTC decoder and inverse string lookup by model itself (no need for extra models and files, all on GPU), always ensemble RGB and binarized input
Added:
- inference backends for TF-Serving and ONNX/TensorRT, differentiate by loaded model type
*`multiprocessing.SpawnProcess` predictor wrapper for models to have commmunication with Tensorflow in a separate subprocess in a task queue with parallel jobs configurable via `--num-jobs` and maximum number of failed jobs via `--halt-fail`
* Keep batch size low enough for processing fitting into common 8GB GPU (with model-dependent batch resizing prepared but not yet active)
* GPU device can be selected manually with `--device`
* Handle image resizing and tiling in GPU as much as possible to avoid overhead of switching between GPU and CPU
* jit-compile and precompile models where possible (non-autosized, non-patched Keras models)
* Fix bugs and homogenize internal labels related to differing labels for early layout and different stages of full layout detection
* Replace `Lambda` layers with `ZeroPadding2D`, improving size and optimizability of models for `eynollah layout`
* Improved training
* Use connected components for loss function
* Integrate with Tensorboard to observe model training progress, including plots and visualizing intermediate evaluation results
* Simplified model usage
* Models can be overridden individually, so any model trained with `eynollah-training` can replace any model in the [distributions on zenodo](https://zenodo.org/records/17727267)
*`--model` is a CLI option of the `eynollah` root CLI now and should point to the same directory for all subcommands
* Improved reading order detection heuristics
* Improved drop capital, marginalia and column detection
* Fixing bugs in polygon handling and image operations
* No more self-intersecting polygons
* Correct rotation implementation, enlarging/shrinking canvas as necessary
* Use actual area of a polygon instead of length of polygon path or first candidate for comparisons
* Improved PAGE-XML serialization
* Annotate column classifier result in `/PcGts/Page/@custom` (Transkribus convention) and `/PcGts/Metadata/Comment` (QURATOR convention)
* Annotate page skew in `/PcGts/Page/@orientation`
* Calculate and annotate confidences as `Coords/@conf` for regions, lines, images and tables
* Massive refactoring and code quality improvement
* deduplication, idiomatic python, clean parallel processing, class reuse, consistent and meaningful naming
**NOTE** We are aware of a possible issue with regards to the cropping of images. It appears that we have not consistenly cropped images for training. This can lead to suboptimal results for cropped images. If you experience quality issues with the `eynollah layout`, try setting the `-ipe/--ignore_page_extraction` option to skip the builtin cropping. We will rectify this in the next trainings.
*`eynollah-training` CLI and docs for training the models, #187, #193, https://github.com/qurator-spk/sbb_pixelwise_segmentation/tree/unifying-training-models
* Drop capitals are now handled separately from their corresponding textline
* Marginals are now divided into left and right. Their reading order is written first for left marginals, then for right marginals, and within each side from top to bottom
* Added a new page extraction model. Instead of bounding boxes, it outputs page contours in the XML file, improving results for skewed pages
* Improved reading order for cases where a textline is segmented into multiple smaller textlines
* For the lightweight version (layout and textline detection), thresholds are now assigned to the artificial class. Users can apply these thresholds to improve detection of isolated textlines and regions. To counteract the drawback of thresholding, the skeleton of the artificial class is used to keep lines as thin as possible (resolved issues #163 and #161)
* the binarization model is added into the models and now binarization of input can be done at the first stage of eynollah's pipline. This option can be turned on by -ib (-input_binary) argument. This is suggested for very dark or bright documents
