Releases: knaw-huc/laypa
1.3.12
Release Notes for Laypa Version 1.3.11-1.3.12
Date: 2024-3-22
Overview
Bug fixes for squarelines and AMP. Duplication detection in preprocessing
New Features
- Duplication detection based on filename: Detect duplicates and where the are coming from. Also useful to see overlap in GT datasets (if names are meaningful)
Enhancements
- Fix issue with color drawing of square lines: Panoptic segmentation with square lines is now supported
- Fix top bottom drawing in: bug made all previously drawn lines part of the top bottom calculation
- Fix typo that prevented AMP during training: See Deprecations and Removals for how to use it
Deprecations and Removals
- AMP is turned off by default: Users experience issues with bfloat16. To prevent multiple people coming with this question the AMP has been turned off. To turn it back on for training and inference use
MODEL.AMP_TRAIN.ENABLED
andMODEL.AMP_TEST.ENABLED
respectively.
Known Issues
- None
Docker Image
The Docker image for version 1.3.12 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.12
Contributors
- @stefanklut: All
Full Changelog: 1.3.10...1.3.12
1.3.10
Release Notes for Laypa Version 1.3.9-1.3.10
Date: 2024-2-2
Overview
Automatic Mixed Precision, squaring of lines, API update, and fixing weird OOM stuff
New Features
- Automatic Mixed Precision: Added optional arguments to use Automatic Mixed Precision (AMP) during training and inference. Current default casting has been set to bfloat16
- Square Lines: Added option to draw square lines. First drawing in lines and retroactively assigning what pixels are part of the beginning and end of the baseline. See Known Issues
Enhancements
- API health point: Check if API has not crashed
- Improvement to resume training method: Only requires a config file now, infers the checkpoint from the config
- OOM Error handeling: Remove traceback from error to prevent OOM from crashing the full program
Deprecations and Removals
- None
Known Issues
- Square Lines for training pano: Currently broken, but already fixed in next release cycle
Docker Image
The Docker image for version 1.3.10 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.10
Contributors
- @TimKoornstra: OOM error and most of the AMP setup/testing
- @stefanklut: The rest
Full Changelog: 1.3.8...1.3.10
1.3.8
Release Notes for Laypa Version 1.3.8
Date: 2024-1-18
Overview
Refactor XMLConverter and OutputPageXML
New Features
- New region config files: New region configs for rekesten (Republic)
Enhancements
- Pull out XMLRegions as seperate class: Rewrite inner workings of XMLRegions
- XMLRegions as input of XMLConverter: Change to use class
- XMLRegions as input of OutputPageXML: Change to use class
- Fix config file: Typo in config made it not work
Deprecations and Removals
- None
Docker Image
The Docker image for version 1.3.8 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.8
Contributors
- @MMaas3: Initial refactor
- @stefanklut: The rest
Full Changelog: 1.3.7...1.3.8
1.3.7
Release Notes for Laypa Version 1.3.7
Date: 2023-12-22
Overview
Fixing UUID and githash issue
New Features
- Crop: Crop image and sem_seg
- Orientation: Change the orientation of an image by order of 90 degrees
- Rectangle around regions: Convert sem_seg predictions to rotated rectangles (for example for the
- Eval model without GT: Look at prediction results without having the GT present
Enhancements
- Better Error handeling: Actual error message that were missing for some errors
- Remove sklearn dependency: Save some space on install
- Better docstrings: Update and add many docstrings
- Use correct affine tranform: use cv2 to match what the coords function was doing
Deprecations and Removals
- None
Docker Image
The Docker image for version 1.3.7 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.7
Contributors
- @MMaas3: Initial rectangle for photos code
- @stefanklut: The rest
Full Changelog: 1.3.6...1.3.7
1.3.6
Release Notes for Laypa Version 1.3.5-1.3.6
Date: 2023-12-11
Overview
Fixing UUID and githash issue
New Features
- Exif: Handle Exif rotation when loading in images
Enhancements
- Fix loading issues Mask2Former: Currently running correctly, with known issue that annotations must have at least one instance
- Saving images issue resolved: Models trained with this bug may need to be changed to "BGR" INPUT
- Fix dataset reloading during validation: Ground truth can now be moved during training, preprocess now links images when not resized
Deprecations and Removals
- Saving images issue resolved: Models trained with this bug may need to be changed to "BGR" INPUT
Docker Image
The Docker image for version 1.3.6 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.6
Contributors
- @stefanklut: Only maintainer currently
Full Changelog: 1.3.4...1.3.6
1.3.4
Release Notes for Laypa Version 1.3.1-1.3.4
Date: 2023-11-22
Overview
Fixing UUID and githash issue
New Features
- Write info to pageXML Take over functionality from loghi-tooling to write UUID and githash to pageXML. TURN OFF
-laypaconfig
in loghitooling. - Add whitelist Whitelist can take multiple joined (by dot) keys (e.g. MODEL.BACKBONE).
To call whitelist add them to the API call usingwhitelist
as argument. For each item whitelisted add a seperate form entry.
Example:-F whitelist=MODEL -F whitelist=INPUT
Forrun.py
use the command line argument--whitelist
or-w
.
Example--whitelist MODEL
or-w MODEL INPUT
for 2 whitelisted items
Enhancements
- Fix githash Get githash from inference code
Deprecations and Removals
- None
Docker Image
The Docker image for version 1.3.4 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.4
Contributors
- @stefanklut: Only maintainer currently
Full Changelog: 1.3.0...1.3.4
1.3.0
Release Notes for Laypa Version 1.3.0
Date: 2023-11-14
Overview
Top bottom split for baselines, and many quality of code improvements
New Features
- Top-Bottom: Split baseline in top bottom, based on pixel location in relation to line segments
- Github Action Black: Auto-formatting on push and pull request
Enhancements
- Pillow supported image formats: Single location for all image formats, retrieved directly from Pillow
- Loading without all methods implemented: When running the code, now able to not have all types supported (top-bottom only does sem-seg) and not crash
- Load in RGB: Change loading to use normal RGB (WARNING: if using old configs that don't specify everything, this might break something. All configs from training output should still work. normal baseline_dataset.yaml will not work with old weights.)
- Speed-up Overwrite check: Don't load entire json to check image size for speedup + Overwrite when changing modes
- Warn for overlapping baselines: During pre-processing
- Change Conversion Method: Other method for converting to numpy for more supported image formats. Solves alpha channel problem for example.
Deprecations and Removals
- Load in RGB: Change loading to use normal RGB (WARNING: if using old configs that don't specify everything, this might break something. All configs from training output should still work. normal baseline_dataset.yaml will not work with old weights.)
- INPUT.SCALING: DeprecationWarning INPUT.SCALING is losing support; please switch to INPUT.SCALING_TRAIN and INPUT.SCALING_TEST
Docker Image
The Docker image for version 1.3.0 can be obtained using the following command:
docker pull loghi/docker.laypa:1.3.0
Contributors
- @stefanklut: Only maintainer currently
Full Changelog: 1.2.10...1.3.0
1.2.10
Release Notes for Laypa Version 1.2.6-1.2.10
Date: 2023-10-27
Overview
Refactor and image loading changes
New Features
- Case insensitive file extensions: Support for (for example) .PNG .JPG
- Direct loading tensors: Load png and jpg directly to tensor (not currently on due to tiff support being required)
Enhancements
- More accurate status codes API: Updated status code to more accurately reflect the error type
- Model loading location: Load during inference not during /predict
- Bugfix loading model during training: Correctly load model to start training from previous step (TRAIN.WEIGHTS)
- Refactor xml_converter: Move critical conversion parts to this file instead of the pageXML format
- Standardize formatting: Following black formatter for python
Deprecations and Removals
- None
Docker Image
The Docker image for version 1.2.10 can be obtained using the following command:
docker pull loghi/docker.laypa:1.2.10
Contributors
- @stefanklut: Only maintainer currently
Full Changelog: 1.2.5...1.2.10
1.2.5
Release Notes for Laypa Version 1.2.4-1.2.5
Date: 2023-10-16
Overview
Minor docker bug and documentation
New Features
- None
Enhancements
- Update Documentation: Add function descriptions and type hints to help people understand the code
- Minor bugfixes: Some missed changes when dealing with configs other than the basic configs
- Fix Docker: Get docker with cuda working again after breaking due to update
Deprecations and Removals
- None
Docker Image
The Docker image for version 1.2.5 can be obtained using the following command:
docker pull loghi/docker.laypa:1.2.5
Contributors
- @stefanklut: Only maintainer currently
Full Changelog: 1.2.3...1.2.5
1.2.3
Release Notes for Laypa Version 1.2.3
Date: 2023-10-03
Overview
This version improves pre-processing speed, adds a scaling option for resizing and reading out NVIDIA Env extraction.
New Features
- Automatic NVIDIA Env extraction: Get info like cudatoolkit version and GPU used. Used for debugging and README.md
- Scaling Resize: Scale input images based on a percentage. Used in both pre-processing, training and inference.
Enhancements
- Faster Pre-processing: If no resize occurs use sym/hard links to save memory and processing time.
Deprecations and Removals
- Pre-process Resize Flag:
PREPROCESS.RESIZE.USE = False
Is currently no longer the method to change resizing. Please usePREPROCESS.RESIZE.RESIZE_MODE = "none"
Docker Image
The Docker image for version 1.2.3 can be obtained using the following command:
docker pull loghi/docker.laypa:1.2.3
Contributors
- @stefanklut: Only maintainer currently