CHANGES.md

Changes

0.9.0

• Comprehensive support for R based scripts (R plugin)
• Change default location of source code to run directory root
• Use .guild directories as default GUILD_HOME
• Support for Python 3.11
• Drop tests for Python 3.6
• Bring scikit-optimize under Guild repo (version 0.9)
• Bug fixes

0.8.2

• Merge command
• Run file manifest to track copied run files
• Guild file schema
• Commands for API level functionality
• Dark mode for TensorBoard
• Support latest versions of scikit-learn
• Support for PowerShell and pwsh on Windows
• Support for debugging via debugpy
• Bug fixes
• Officially drop support for Python 2

0.8.1

NOTE: If you use guild's autocomplete functionality, you will need to re-install the completion scripts, as they have changed. It should be sufficient to run guild completion --install.

• Add Pydantic typing information to the core data structure classes. This is used to generate a schema for the guildfile.
• Add python-based autocompletion implementation. Bash behavior has not changed, but more completions should now work on zsh and fish.
• Add support for BooleanOptionalAction from argparse in Python 3.9+
• Removes deprecated flags support
• Update click dependency to >=8.1

Fixes:

• Get tests working with Python 3.10
• Fix issue with guild init where python version was ignored
• Rename flags() attribute of Run class to guild_flags(), to avoid conflict with pandas >=1.2

0.6.5

• Remove TensorFlow requirement
• Convenience option to view and open run output
• Save platform information for runs
• Check latest version of Guild AI in check command
• Remove dependency on TensorFlow
• Sensible default run labels (specifies non-default flag vals)
• Simplify output scalar patterns with \key and \value regex aliases
• Batches are no longer included in the TensorBoard view by default

Fixes:

• Performance of compare with resource directories containing large numbers of files
• Bug in reading flags from Python scripts

0.6.4

• TensorFlow 2 support (beta)
• Refactor publish implementation
  • Publish series of user-facing run files (e.g. run.yml, flags.yml, scalars.csv, etc.)
  • Don't publish files by default (can explicitly publish using --files option)
  • Complete include and exclude support for selecting files to publish
  • Include run output in default report
• select-min and select-max patterns for reducing source file selection to a min and max version respectively
• User script exceptions are shown with Guild stack layers removed to reduce noise (this behavior can be disabled for additional debugging)
• Support for port and connect-timeout SSH remote attributes

Fixes:

• Import error when using guild.ipy (missing click module)
• Cleanup use of labels for batch trials

0.6.3

• Early release support for publishing runs
• Early release Notebook support (guild.ipy module)
• Renamed source to snapshot-source to disambiguate from resource source config
• Simplified snapshot source config
• Safe guards for default source snapshots (can be overridden by adding snapshot-source config)
  • Skip files larger than 1M
  • Don't copy more than 100 files

0.6.2

• Improved scheme for capturing script output as scalars
  • Two-group captures used for key/value logging
  • Named group captures
• Show all flags and scalars by default in compare
• Show scalar values in runs info (requireds --scalars option) (previously only scalar names were shown)

Fixes:

• Distribution dependency on scikit-optimize

0.6.1

• Windows support for Python 3.5, 3.6, 3.7

Fixes:

• Fix import of boolean flags on Python 3 (report by @OliverRichter)
• Skip all dot directories during source snapshots
• Skip archive diretories during source snapshots

0.6.0

This is the baseline release of Guild AI.

Major features:

• Run, track, and compare experiments
• Hyperparameter optimization using grid search, random search, and Bayesian optimzation
• Automate model operations and workflows using Guild files
• TensorBoard integration
• Remote training, backups, and restore