v0.8.398

• adding the new validation tests
• Merge branch 'main' into validation_support
• Pin awswrangler to its Python/pandas engine
• Merge branch 'main' into validation_support
• adding a reconnect loop to the redis tunnel
• Merge branch 'main' into validation_support
• improve the comment about the sagemaker-core version pin
• bumping version of workbench for all images
• validation-set support: de-overload sample_weight into explicit validation/exclude row roles

v0.8.397

• something scary happened in sagemaker-core==2.13; so pinning down core and related dependencies
• adding some logic to handle partial creation of FeatureSet
• new planning documentation
• adding a workbench.training import path for code/utils that are used during the training phase for model scripts
• including the templates in the flake8/black coverage
• adding some hyperparameter logic based on what's set
• adding a chemprop model with some tuning knobs changed as an example
• adding a chemprop model with some tuning knobs changed as an example
• removing old pxr/storage scripts
• updating the UQ docs
• updating the 3D docs, files, and tests to new naming and functionality
• updating our chemprop models docs
• bumping version of workbench for all images

v0.8.396

• adding 0.8.396 release notes
• Remove MVE experimental files (negative result): phase1 MVE model, seed-sweep + analysis scripts, pipelines.json entr
• Remove MVE integration — negative result: NLL compresses ensemble epistemic signal (seed sweep, all 3 seeds); keep chemprop 2.2.4 LR/FFN knobs
• adding aleatoric std column to our saved inference output
• bumping version of workbench for all images

v0.8.395

• Add chemprop MVE (mean-variance) head option that feeds per-molecule aleatoric std into UQ v1 as an error-model feature, plus a phase-1 PXR MVE model for held-out evaluation
• updating to chemprop 2.2.4
• bumping version of workbench for all images

v0.8.394

• add a test for exact file names used in the ml_pipeline launcher
• logic for running only one script when a full filename is given
• improve the chunk progress display
• only scanning the files in the PR/push
• switching over to local binary
• change how plugins are located
• adding a workflow for leak detection
• adding a precommit config for leak detection
• make sure all async children have their max_instances set up
• renaming some of the endpoints
• renaming some of the endpoints
• new feature endpoint for 2d (keep salts) + 3d
• new dynamic batch size computations for meta endpoints
• updating our Open ADMET PXR blog
• updating our Open ADMET PXR pipelines/models
• adding a phase2 model for open admet pxr
• flake8/linter fixes
• adding multi-task models for PXR experiments
• adding chemeleon foundation models
• bumping version of workbench for all images

v0.8.393

• improve the model copy() docstring
• moving copy_model_artifacts() into model_utils
• add a TBLITE_AVAILABLE on the test logic
• flake8/linter fixes
• fix the S3 prefix when creating model copies
• bumping version of workbench for all images

v0.8.392

• flake8/linter fixes
• adding a test for our 3D v2 descriptors
• adding inputs to our pipelines.json for PXR
• removing an unused arg/param in the 3D v2 descriptors
• updating the PXR models with 3D data
• adding a version 2 of our 3D descriptors for testing
• small change to log messages
• adding the 3D descriptor pipeline details SVG
• updating the 3D conformer details diagram
• bumping version of workbench for all images
• updating the 3D conformer tier in the documentation

v0.8.391

• new timing and conformer tier logic
• bumping version of workbench for all images

v0.8.390

• always fire to warm-up request since it simplifies code logic
• adding a PER_CONFORMER_TIMEOUT_S for more complicated compounds
• removing all of the docs/scripts for the deprecated 'fast' 3D endpoint
• preload the InferenceCache class into the REPL
• bumping version of workbench for all images

v0.8.389

• Merge branch 'main' of github.com:SuperCowPowers/workbench
• reorganizing the public_data reference compound upload
• reorganizing the public_data reference compound upload
• don't use params.useSmallRingTorsions = True it was causing embedding failures
• parallelizing the warmup calls
• adding a warm up test
• adding warmer logic for async endpoint so that we don't hit timeouts while AWS spins up from 0
• bumping version of workbench for all images