- Add flake8 to the CI.
- Use the metalearning weights when all actions have equals UCB.
- Use the mean adtm values as probabilities for the patterns.
- Create a compressed Meta-learning DB.
- Add support for LUPI tasks.
- Fix CI.
- Send the path (to save model weights) as hyperparameter.
- Remove warning about the absolute paths.
- Show error from scoring process.
- Use raw datasets for the examples.
- Add datasets for the examples.
- Add observations to the installation procedure.
- Update documentation.
- Unlock d3m dependencies.
- Update documentation.
- Add 'resource_folder' parameter to the API.
- Remove unused data reader module.
- Add more Jupyter Notebook examples.
- Add alpha-containers package.
- Restructure repository and implement a new API.
- Read version from file.
- Update instructions for releases.
- Move dependencies from setup.py to requirements.txt.
- Use D3MPORT env variable.
- Added instructions for releases.
- Relax requirements.
- Disable reuse port in GRPC.
- Implemented the PyPI version (lightweight)
- Added support to to blacklist and whitelist primitives
- Added support to save and load pipelines.
- Use timeout_run parameter to score pipelines during search.
- Use the AutoML RPC utils for encoding GRPC pipelines.
- Implemented the method SaveFittedSolution of the AutoML RPC API.
- Restructure packages (rename d3m_ta2_nyu to alphad3m).
- Expose outputs of each step within the pipeline.
- Build automatically the grammar from the metalearning database.
- Calculate primitive correlations.
- Added scripts for running on SLURM/Singularity.
- Added encoders (text, datetime, etc.) to the grammar.
- Added timeout for pipeline execution (during search).
- Added support to expouse outputs of the pipeline steps.
- Added support for ROC AUC metric.
- Renamed repository to AlphaD3M.
- Added support for video data type.
- Improved support for semi-supervised task through SemisupervisedClassificationBuilder class.
- Updated license to Apache-2.0.
- Added support for clustering problems.
- Created NN inputs for AlphaD3M from the metalearningDB #46 !49
- Changed the structure of the preprocessing module. Added text and datetime encoders.
- Updated to core package v2020.5.18 and TA2-TA3 API v2020.6.2.
Submission for Winter evaluation.
- Added data profiler to the workflow. #39 !47
- Added support for LUPI problems.
- Added encoders to the search by AlphaD3M.
- Updated to core package v2020.1.9 and TA2-TA3 API v2020.2.11.
Submission for December dry-run.
- Added different tasks to the grammar. #35 !42
- Updated sampling strategy. #30 #32
- Added templates to external process.
- Updated to core package v2019.11.10 and TA2-TA3 API v2020.12.4.
- Changed internal versioning to CalVer format.
Submission for June dry-run.
- Added standard Reference Runtime to execute pipelines
- Added data sampling strategies and priorization of some D3M primitives
- Added
RANK
metric (and correspondingRANKING
evaluation method) for TA2-only evaluation - Added
rank_solutions_limit
parameter inSearchSolutions
which allows request both searching and ranking at the same time - Updated TA3-TA2 API functions: ListPrimitivesRequest, SearchSolutions and ScoreSolution
- Updated to core package v2019.6.7 and TA2-TA3 API v2019.6.11
Re-submission for 2018 Summer evaluation after Gov team mixup on TA1 library freeze
- Finish implementing gRPC server (which we were initially planning to do before TA2 and TA3 deadlines)
- Use correct base image, mandated by Gov
- Added an 8-minute timeout to
ScoreJob
(some primitives freeze) - Only report scored pipelines to TA3, don't inform them of created-not-yet-scored (or broken) pipelines
First submission for 2018 Summer evaluation (original deadline).
- Build from common
jpl/docker_images/complete
images - Use
d3m
package to load dataset, remove MIT-LL'sd3mds.py
- Added
eval.sh
entrypoint to support Data Machine's eval protocol - Updated gRPC to v2018.7.7
- Training/testing is now independent of
Session
, which only handles searching - Add a timeout on AlphaD3M
- Do tuning after top pipelines have been trained and written out (then train the tuned pipelines)
- Use
KFoldDatasetSplit
primitive to do cross-validation splits
- Added AlphaD3M pipeline generation
- Enabled huperparameter tuning with SMAC
- Added the
Job
class for the run queue - Introduced own multiprocessing code using sockets and avoiding fork issues
Bug fixes.
Bug fixes.
Bug fixes.
Version submitted to NIST for 2018 January evaluation.
- Added hyperparameter tuning with SMAC. Disabled, does not work
- Raise the number of pipelines by using one of 3 imputers, one of 2 encoder
- Updated gRPC to v2017.12.20
January dry-run version.
- Removed VisTrails
- Moved from Python 2.7 to Python 3.6
- Renamed package from
d3m_ta2_vistrails
tod3m_ta2_nyu
- Use
d3mds.py
from MIT-LL to load dataset - Use some D3M primitives, in addition to "native" scikit-learn:
KNNImputation
andEncoder
from ISI's dsbox
- Added CI
- Updated gRPC to v2017.10.10
Version submitted to NIST for 2017 Fall TA3 evaluation.
- Improvement to data-reading code
- Create directories
Version submitted to NIST for 2017 Fall TA2 evaluation.
- Using gRPC protocol v2017.9.11
- Custom data-reading code, identifies column types, does PCA for image data
Start of project, using VisTrails for workflow representation and execution.