Skip to content

Latest commit

 

History

History
75 lines (75 loc) · 4.77 KB

CHANGELOG.org

File metadata and controls

75 lines (75 loc) · 4.77 KB
Raw

Change-log

2.0.a.7 [2020-12-13 Sun]

  • Minor fix with filtering version in single cli
  • Delayed initialization of datasets

2.0.a.6 [2020-04-12 Sun]

  • Moved the to mulitiprocessing from subprocess
  • Single mode now can test all versions of a single experiment
  • Implemetation of the iterator which breaks on test mode and Removed the `steps` from the main pipeline process
  • Added more flexibility to how datasets are loaded and moved Datasets to utils
    • train_dataset_file_path can now be empty
    • validation_data_load function as a new param
  • Moved the execute_experiment function to Process
  • Version loaded also shows progress in the form of [/]
  • Renamed dataset_file_path to data_asset in utils.Dataset
  • Log dataloader parameters if DataloaderWrapper class is used for the dataloader
  • Added API function to run experiment with instantiated experiment object
  • Method in Versions to add versions from another Versions object

2.0.a.4 [2019-07-24 Wed]

  • Added function to load previously executed experiment.
  • The experiments are executed with the cwd set to the `experiments_dir`.
  • Added base implementation for pytroch.
  • Added a post_execution_hook, which operates the same as the pre_execution_hook and run after all train and evaluate loops are over.
  • Added current_version, experiment_dir and dataloader to be experiments properties
  • Moved to using multiprocessing from using subprocess
  • mlpipeline in single mode allows testing all versions of an experiment
  • Renamed cli script to ‘_cli’
  • Implemented iterator class which can be used to breaking during test mode
  • Since the above was one of the primary reason for the `steps` calculations, removed the `step` calculations from the main pipeline
  • Added more flexibility to how data is being loaded.
    • train_datasets_file_path can be empty
    • new parameter validation_data_load_function
    • If the input to the train or evaluation loop is None, the pipeline still attempts to execute the methods. If both train and test inputs are None, the evaluate_loop will be executed only once. If train input is None while test input is not None, evaluate_loop will execute twice: once with None as the input, and the second time with the test input.
  • None of the methods need to be implemented. If any method is not implemented, it’ll simply log a message instead of throwing a tantrum.
  • Re-organized the modules and functions
  • Bumped minimum version of python to 3.5.5 to deal with the requirement of pandas
  • Bug fixes:
    • Moved each version of an experiment to be executed in a separate sub-process
    • Fixed bug with scheme and windows path for mlflow
    • Fixed bug with cli not completely migrated to using process abstraction
    • Ensured change to cwd when using api.get_experiment and exposed dataloader to the returned experiment

2.0.a.3 [2019-07-16]

  • Several methods from the base classes are not required to be implemented. They are:
    • Experiment.setup_model
    • Experiment.pre_execution_hook
    • Experiment.eval_loop
    • Dataloader.get_test_sample_count
    • Dataloader.get_train_sample_count
  • dataloader now can be None in a version.
  • Added DataLoadercallablewrapper class to base
  • Added is_no_log
  • Made mlflow as an install_dependency
  • mlflow now can be used in test mode as well. The tracking uri is set to the temp dir in test mode.
  • User now can set the mlflow tracking uri
  • Migrated cli interface to click and fixed bugs with SDK

2.0.a.2 [2019-07-10]

  • Major refactor: helper module renamed as base
  • Executed experiments in _pipeline_subprocess.py uses a class instead of dictionary
  • Added export mode
  • Added python API to execute mlpipeline
  • Added parameter agent to mlpipeline.utils.log
  • Fixed bugs as a result of new mlflow API
  • Moved loading scripts to separate function in mlpipeline.utils

1.1a3.post8 [2019-03-04]

  • Major refactor: renamed model to experiment.
  • Functions to track and copy files imported by the experiment scripts.
  • All files imported by an experiment script and the experiment script itself will be copied to the `experiment_dir` directory
  • Removed the use_history feature for now.
    • It seems to be buggy.
  • Mlflow integration.
    • If mlflow exists and use_mlflow is not set to `False`:
      • All mlflow related data will be stored in the outputs/mlrun directory
      • Each experiment script is considered a mlflow-experiment.
      • Each version executed form the experiment is considered a mlflow-run of the respective experiment.
      • The Matric.log_matrics will log all the metrics to mlflow also.
      • All the parameters passed through the version will be logged as parameters by mlflow.