Skip to content

Version Changes v.0.6.3

Jump to bottom
Mikko Kotila edited this page Aug 25, 2019 · 5 revisions

Overview

**DATE : ** 25 of August, 2019

This version change focuses on fixing issues that were introduced in the new major version release 0.6. The aim have been to get the new LTS onto pypi so everyone can start benefiting from the new powerful features.

Summary

  • Many new optimization strategies added
  • Experimental Torch support is added
  • Consistency across all the AutoML features in Autom8
  • Many issue fixes

Housekeeping

  • Docs are now comprehensive, and consistent with the >0.6 Talos API.
  • Scan() docstrings are up-to-date and cleaned up
  • Many other docstrings are cleaned up
  • Docstrings will take the same format in the future as markup table to give consistency with documentations
  • Removed redundant files from the repo
  • Tests are updated in regards to the changes but not yet all new features
  • Added "edit on github" link to docs
  • Added free text search to docs
  • Added some styling to docs
  • Added analytics to docs
  • Added "copy to clipboard" to code snippets in docs
  • Improved PR template
  • Improved issue template
  • Created feature request template
  • Created bugs template
  • Changed the deploy test so local folder does not end up with garbage

What's new

  • Multi-output models are not fully supported #154
  • Reporting() now also accepts Analyze() as its command
  • Reporting() no longer has 'val_acc' as default value in any of the properties
  • early_stopper() now has 'lazy' mode with slight tweaks to other modes
  • early_stopper() no longer expects epochs value for custom settings
  • There is now a new sub-module autom8 inside which several AutoML features live
  • AutoParams automatically generates a parameter dictionary and streamlines its manipulation before experiment
  • AutoModel automatically creates a input model for Scan() which is fully wired for use with AutoParams or other experiment with comprehensive search
  • AutoScan leverage AutoParams and AutoModel to reduce the whole experiment into a single line of code
  • AutoPredict takes the results of Scan (or AutoScan), picks best model candidates, evaluates the candidates, picks the winner, and makes predictions with it on input data
  • Added local_strategy to reduction strategies, which allows making changes to the parameter space from local system while the experiment is running
  • Added pearson and kendall reduction strategies
  • Streamlined the way custom strategies can be added
  • Completely rebuilt correlation strategy, including the underlying statistical approach
  • Added a helper function cols_to_multilabel for custom reducers
  • Added a new generator SequenceGenerator
  • Added talos.utils.ExperimentLogCallback which allows storing epoch-by-epoch training data on the local machine during the experiment (implements the request in #153)
  • experiment_name is now compulsory
  • It's now possible to control the experiment during the experiment and receive live updates on progress Addresses #207 and prepare for browser based "command center"
  • Added 'trees' reduction strategy
  • Added 'forrest' reduction strategy
  • added scan_utils.py as a home for helper functions for /scan
  • man-machine strategy is invoked through reduction_method='gamify' in Scan()
  • Checks and updates a parameter map in the experiment folder each permutation
  • If parameter value status is changed to 'inactive' in the .json locally, then reduction will be applied (all permutations with that value will be removed)
  • Related with this, check out http://github.com/autonomio/gamify ... an add-on for live monitoring of experiments and analysis of experiment results, with coming features for controlling the experiment as well
  • Related with #343 it's now possible to avoid saving model weights in scan_object, which might be desirable for very long runs with very large networks, due to the memory cost of keeping the weights throughout the experiment.
  • max_param_values is now optional in AutoScan and instead created the issue to handle the underlying problem properly #367
  • Added experimental support for Torch
  • Added a couple of initializers to AutoParams

What's fixed

  • Removed the default 'val_acc' so instead have to explicitly state metric when deploying (fixes #283)
  • Fixed a bug related with the case where x is not 2d
  • Deploy now uses the right numpy array property (fixes #348)
  • Restore now has as default allow_pickle=True (fixes #351)
  • Fixed a small bug in AutoParams where choosing network=False resulted in 'dense' to be split into characters
  • Also fixes #367 (setting 4 per paramater as the default for automatic mode)
  • Small issues on AutoModel() were fixed
Clone this wiki locally