HISTORY.rst

History

Unreleased

2.2.1 (2020-10-14)

Changes

• N/A

Fixes

• Fix incorrect retrieval of latest Dataset due to invalid search criteria of FINISHED status. Both lower and upper naming variants will be employed for backward compatibility of existing Dataset entries.

2.2.0 (2020-10-01)

Changes

• Provide crop_mode parameter to BatchCreation process that specifies whether to generate extend (pad crops to have squares), raw (original minimal rectangle) or reduce (clip bbox to make a square). When crop_fixed_size is not provided (has priority), crop_mode indicates how to generate the patches. Default is extend (most typical use case: square bbox of various sizes to be handled by preprocessing transforms). Previously, the minimal rectangle bounding box fitted to contain the annotated feature were directly used to generate non-squared patch crops. The dimensions are now adjusted to specification [GEOIM-238].
• All Dataset now only contain one type of patch crop (fixed, extend, fixed, reduce) according to the parameters provided during BatchCreation request. Patches of fixed size are no longer added on top of the raw crops in the same Dataset.
• Loads all crops of each patch within the Dataset as individual sample for the data loader. Crops where previously assumed to always be a list of a single item. They will support multiple if more are provided.
• Provide map of masked image labels to samples generated by the test data loader for Segmentation task.
• Add summary result reporting function extended by specialized ones for Classification and Segmentation tasks.
• Forward dontcare configuration parameters to the data loader for proper sample generation.
• Add more validation of input model task to have expected class IDs format matching taxonomy classes to map against.
• Add support of model upload using Google Drive shared file link.
• Enforce geoimagenet_ml.ml.data_trainer_device configuration to use the explicit device available. Don't rely on possibly unavailable/invalid device inferred from model checkpoint uploaded by external user.
• Update requirements to crim-ca/gin-model-repo (1.0.1) and thelper (v0.6.2) to employ latest updates at this time.

Fixes

• Fix creation of crop masks as UINT8 black/white PNG images [GEOIM-337].
• Fix test runner for Classification task with adjustments to job update and other parameters incorrectly passed down during sample generation by the data loader.
• Revert erroneously removed implementation of data loader for Classification (was overridden by Segmentation implementation in 2.0.0).
• Pin celery<5 that breaks imports.

2.1.0 (2020-09-28)

Changes

• add crim-ca/gin-model-repo (1.0.0) to dependencies to facilitate import of corresponding utilities and model implementations

2.0.0 (2020-09-15)

Changes

• add scripts/update_model_classes.py to update taxonomy class indices of model task in checkpoint (CLI script callable after install with gin_ml_update_model_classes)
• add scripts/run_model_tester.py to execute ProcessRunnerModelTester process offline (CLI script callable after install with gin_ml_run_model_tester)
• add memory implementations of database stores for testing and scripts
• add support of Segmentation networks inference with corresponding dataset loader, metrics and tasks
• update thelper>=0.5.0 requirement for latest changes and features

Fixes

• fix validation of model task if already parsed as instance of thelper.tasks.Task
• fix incorrect crop_fixed_size input description of ProcessRunnerBatchCreator
• fix Classification task execution of ProcessRunnerModelTester with updated thelper definitions
• fix saved image crop mask of annotated features in dataset

1.5.0 (2020-06-11)

Changes

• add support of MultiPolygon feature if it is defined as a single set of Polygon during batch-creation
• drop Sentry events in case of HTTP Client errors (4xx) to avoid over reporting expected responses by API

Fixes

• filter directory walker file types to tif only for retrieval of rasters within batch-creation process runner

1.4.1 (2019-08-08)

Changes

• remove git clone of thelper in Jenkinsfile as it is installed with requirements pinned version from PyPI

Fixes

• fix incorrect descriptions returned for GET requests of job logs and exceptions
• fix key error when parsing model task defined as literal string [GEOIM-245, GEOIM-247]
• pin gdal==2.4.2 as gdal==3.0.1 fails on import (OSGeo/gdal#1759)

1.4.0 (2019-08-06)

Changes

• add error logging in case of request generating an exception (voluntary or by execution problem)
• change gdal requirements and imports to avoid osgeo variant breaking too easily
• support older thelper model checkpoint definitions using string parameters instead of JSON [GEOIM-241]
• update to thelper>=0.3.7, help better support old format for model task definitions [GEOIM-247]
• disable Jenkins online tests execution (download/upload model) as they now require login [GEOIM-180]

Fixes

• avoid key error of WPS-like process properties cleanup during initialization [GEOIM-225]
• better filter and group common job error message to Sentry [GEOIM-227]
• update test model checkpoint for unittest execution with more relevant thelper configuration [GEOIM-180]

1.3.0 (2019-06-18)

Changes

• add additional metrics (Top-1, Top-5) in process model-tester execution results (GEOIM-163)
• add fetching of taxonomy definition during batch-creation process execution from input URL (GEOIM-161)
• save retrieved taxonomy definition to generated dataset batch (GEOIM-162)
• enforce thelper>=0.3.1 to use new operation modes and back compatibility import methods

• add model checkpoint validation checks and raise ModelValidationError (HTTP Forbidden [403]) if invalid

  • disallow loading a model task defined as literal string, must be a JSON definition of parameters
  • disallow loading a model task not matched within existing job mapping (processing steps must be fully defined)
  • disallow loading a model task without minimal configuration parameters required during model-tester execution

• add named key variables to help understand corresponding items across dictionary definitions and generated results
• add more detail to produced results of model-tester process (GEOIM-163)

Fixes

• fix incorrectly returned message from POST /processes/{id}/jobs requests in case of error

1.2.0 (2019-06-11)

Changes

• add BatchTestPatchesDatasetLoader implementation that loads a generated dataset of test patches from process batch-creation execution to evaluate them against a registered model using process model-tester.
• add geoimagenet_ml.ml.jobs_path configuration setting to indicate where process job logging should be stored. (mostly during process model-tester execution)
• adjust model and dataset definitions to employ the same task until resolved (GEOIM-153)

Fixes

• fix bug generated during model-tester process execution attempting to update Dataset parameters (GEOIM-149)

1.1.2 (2019-06-06)

• add variable GEOIMAGENET_ML_SENTRY_SERVER_NAME to allow overriding server_name value reported by sentry.
• fix bug caused by invalid sub-item type checker of job exception field (GEOIM-146).

1.1.1 (2019-06-05)

• fix bug caused by invalid list/tuple concatenation in job exception field (GEOIM-145).

1.1.0 (2019-05-16)

Changes

• fetching latest job for processes where limit_single_job=False will not raise 500 if job count >1
• fetching current job for processes where limit_single_job=False will raise 403 because of multiple matches
• fetching current job for processes where limit_single_job=True raises 404 with more appropriate message
• change job mark_started and mark_finished methods to update_started_datetime and update_finished_datetime respectively to be more specific since they do not actually set the status field
• use enum for current and latest keywords
• add additional input format validation during job submission
• add filtering of job search with multiple STATUS and/or CATEGORY simultaneously
• add tests for current and latest jobs use cases
• add tests for job submission input type validation
• update bump version Makefile targets and config
• enforce typing of enum string sub-type and unique constraint

Fixes

• fix process limit_single_job field incorrectly set in database
• fix process reference field incorrectly set in database
• fix double dot (. .) string ending not correctly cleaned up for response, notably process abstract field (requires process recreation or update if already inserted in the database)
• fix typing of ExtendedEnumMeta.get() method return value for expected corresponding enums

1.0.0

Changes

• add visibility update PUT requests for Job and Model (GEOIM-137)
• add strong and enforced input validation of datatype parameters
• upgrade db to version "4", loading previous objects could cause errors (input validation failures)
• add more unittests for input validation
• add test for new visibility routes (GEOIM-137)
• add test for job submission
• add sentry-sdk integration (GEOIM-118)

Fixes

• fix returned body response from job submission to match rest of API format
• fix multiple API schema definitions

0.8.0

Changes

• add request to store corresponding user-id to db if specified with MAGPIE_USER_URL
• add statistics and action tracking of API requests
• add user creating a new dataset, model, job, process
• add started timestamp for jobs not immediately running (accepted but pending), duration based on it
• restructure enum components used across the project

Fixes

• fix rare race condition of job update caused by updated job details not retrieved from db
• fix incorrectly saved datetime as string in db
• fix API schemas and drop unused items

0.7.1

Changes

• add pip check on install to ensure all package requirements/dependencies are met recursively

Fixes

• fix supervisor path reference to source
• fix db invalid index reference

0.7.0

Changes

• rebase source directory from src to geoimagenet_ml to solve installation/debug issues
• add more validation of job inputs
• add and fix utility make targets

Fixes

• fix gdal package and unresolved symbol error
• fix typing and general code formatting
• fix and complete batch-creation job execution

0.6.x

Changes

• Redefine most of the process creation procedure.
• Batch of patches creation process
• Model testing process
• Automatically create default processes on start if not available in db.
• Add dataset download route.

Fixes

• Fix typing and validations.

0.5.x

Changes

• More refactoring and fixes for functional ML on server.

0.4.x

Changes

• Full refactoring of project directories and imports.

0.3.x

Changes

• Setup databases, datasets, models, processes and other interfaces with REST API.
• Setup API schemas for documentation.

0.2.x

Changes

• Switch between mongodb/postgres databases (postgres schemas not all supported)

0.1.x

• Initial release.