Version 3.5.10

Minor:

• Disable mixed precision calls
• Move data IO management to biapy/data and create imread/imwrite functions to avoid using skimage.io as they are deprecating these functions
• Change timm.optim call to avoid warning of future deprecation
• BioImage Model Zoo related (BMZ) changes:
  • Add function to autogenerate a documentation.md
  • Move BMZ related functions to bmz_utils.py
  • Add test_model at the end of BMZ model creation with a larger tolerance than the default so the differences due to casting are allowed
  • Create cover extracting a patch containing mask information
  • Add argument to provide the dataset id when exporting a BMZ model
  • Ensure weights_only=True during checkpoint loading stage when building a BMZ model

Bugs fixed:

• Checkpoint load on DDP right after training fixed
• Update installation restricting torchmetric version to avoid an issue in classification workflow

Full Changelog: v3.5.9...v3.5.10

Version 3.5.9

Major:

• Add synapse segmentation options for instance segmentation (in experimental state)

Minor:

• Add script to convert instance segmentation datasets into detection workflow format
• Print a better message when shapes does not match between samples
• Some variables for detection has been modified and now don't need to be set per class values:
  • Change TEST.POST_PROCESSING.REMOVE_CLOSE_POINTS_RADIUS default value to 0
  • Change TEST.POST_PROCESSING.DET_WATERSHED_FIRST_DILATION default value to [-1,-1]
  • Change TEST.DET_MIN_TH_TO_BE_PEAK default value to 0.2
  • Change TEST.DET_TOLERANCE default value to 10
• Add instance segmentation multihead test in run_checks.py
• Update convert_old_model_cfg_to_current_version function to cover new changes

Bugs fixed:

• Handle multiple data within Zarr/H5 during test
• Delete channel restriction when ensuring 3D shape (convert_instance_data_to_detection.py)
• Fix class prediction to the points in detection
• Fix error with diplib package
• Fix issue between TRAIN.PATIENCE and TRAIN.LR_SCHEDULER.REDUCEONPLATEAU_PATIENCE
• Solve issues with data type during detection watershed so the instance properties can be measured with diplib as it does not support int64 data type
• Fix issue when multiple raw images (lightmycells case) were provided
• Fix issue with BMZ model exportation
• Solve issues with run_checks.py due to recent changes. Now it is correctly reporting when a test crashes as it will crash too.

Full Changelog: v3.5.8...v3.5.9

Version 3.5.8

Bugs fixed:

• Fix Torchvision calls for semantic seg, detection and instance segmentation workflows

Full Changelog: v3.5.7...v3.5.8

Version 3.5.7

Major:

• Rebuilt CartoCell tutorial organization and update notebooks.
• Update templates to follow the same configuration as in the notebooks, which achieve good results in the example datasets.

Minor:

• Improve robustness loading 3D images
• Make SurfaceArea only requested in 3D images
• Update example dataset paths to raw and label in most cases to be consistent

Bugs fixed:

• Fix bug during TEST.REDUCE_MEMORY
• Fix errors while loading H5 nested data
• Solve bug when loading Zarr/H5 files into memory for training
• Fix missing import in some workflows
• Changes in instance segmentation's statistic calculation:
  • Add diplib library as a dependency to calculate surface area more precisely and enable elongation for 3D which is P2A in diplib
  • Correct centroid coordinates
  • Make SurfaceArea only requested in 3D images to accelerate the process
• Fix bug in the filtering while predicting by chunks

Full Changelog: v3.5.6...v3.5.7

Version 3.5.6

Major:

• Add configuration file backward-compatibility
• Add U-NeXt V2 model
• Actions added:
  • Add check_code_consistency.yml action to test code consistency (every week)
  • Add upload_biapy_to_pypi.yml to automatically create a PyPI package (when a new release is created)
  • Add create_release_container.yml file to automatically create and update docker containers to Dockerhub (when a new release is created)

Minor:

• Update BMZ model creation and compatibility:
  • Add cover creation and create environment.yaml to be packaged in the BMZ model
  • Add sigmoid activation as BMZ postprocessing so we are more compatible
  • Extract just the pytorch_state_dict from the checkpoint when creating BMZ package
  • Save correct input/output (prediction) for BMZ package
  • Move to bioimageio.core==0.7.0
  • Change slightly the normalization so it can match the one done in BMZ

Bugs fixed:

• Fix BMZ model compatibility checks
• Update notebooks to avoid BMZ error when fields are None
• Fix bug on BMZ zip creation in the notebooks
• Fix missing letter 'S' in configuration variable 'SIGNS'.
• Disabling percentile clipping as that is not done by default in BMZ's scale_range normalization

Full Changelog: v3.5.5...v3.5.6

Version 3.5.5

Major:

• Add backward compatibility loading checkpoint

Minor:

• Change TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.STAT to TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.STATS
• Only check lr scheduler when train in enabled

Bugs fixed:

• Fix a bug in DATA.FILTER_BY_IMAGE
• Update 3D_cell_detection_zarr_tutorial.yaml with new configuration

Full Changelog: v3.5.4...v3.5.5

Version 3.5.4

Major:

• Add BMZ exportation through configuration

Minor:

• Set automatically BMZ path and change it to PATHS.BMZ_EXPORT_PATH

Bugs fixed:

• Fix minor bug when loading model checkpoint
• Fix small bug in semantic seg. multiclass jaccard calculation

Full Changelog: v3.5.3...v3.5.4

Version 3.5.3

Major:

• Update BMZ model check to support more models and increase it's robustness.

Minor:

• Add class extraction for semantic seg. BMZ models.
• Adapt instance segmentation channels to a default value depending when loading BMZ models.
• Change LOAD_MODEL_FROM_CHECKPOINT default value to True.
• Increase UNETR building process robustness

Bugs fixed:

• Fix bug when filtering by entire images.
• Prevent top-5-accuracy when classes are less than 5 in classification workflow.
• Fix bug in single data generator used in classification and SSL workflows.
• Allow BMZ/Torchvision models override completely configuration with the variables they are imposing by making update_dependencies() config function more generic.
• Force entire image filtering when DATA.EXTRACT_RANDOM_PATCH is enabled.

Full Changelog: v3.5.2...v3.5.3

Version 3.5.2

Major:

• Add 'resunet_se' to I2I workflow
• Extend BMZ model support
• Remove DATA.TRAIN.MINIMUM_FOREGROUND_PER. Now for training, validation and test a sample filtering can be made with DATA.TRAIN.FILTER_SAMPLES, DATA.VAL.FILTER_SAMPLES and DATA.TEST.FILTER_SAMPLES respectively.

Minor:

• Add MODEL.LOAD_MODEL_FROM_CHECKPOINT variable
• DATA.PREPROCESS.MEDIAN_BLUR.FOOTPRINT changed to DATA.PREPROCESS.MEDIAN_BLUR.KERNEL_SIZE
• Add robust semantic mask check using DATA.*.CHECK_DATA
• Divide BMZ model check into two functions so they can be reused easily by the GUI
• Change BMZ COLLECTION_URL to a new version of it

Bugs fixed:

• Fix bug in detection workflow when predicting with Zarr/H5 by chunks
• Minor fix during instance training data creation using Zarr
• Correct minor errors during BMZ model import/export

Full Changelog: v3.5.1...v3.5.2

Version 3.5.1

Major:

• Add GRN, ConvNeXtBlock_V2 and UpConvNeXtBlock_V2 blocks

Minor:

• Change CENTRAL_POINT_DILATION from int to list
• Add support for fixed_zero_mean_unit_variance preprocessing for BMZ models
• Adapt BMZ model check function to work properly with models in 0.4 and 0.5 version

Bugs fixed:

• Bug in instance segmentation using only C channel
• Correct tags in BMZ model creation
• Adapt BMZ model check function to work properly with models in 0.4 and 0.5 version

Full Changelog: v3.5.0...v3.5.1