Release 0.7.3

Enhancements

Dask based scheduler for parallel runs

Guild now supports parallel runs through a Dask based scheduler. For more information, see Parallel processing with Dask scheduler.

Support for using extra-batch runs for optimization trials (prev-trials optimization flag)

Guild now supports prev-trials flag for all sequential optimizer operations (e.g. gp, forest, and gbrt). This value can be batch (default), sourcecode (use trials that match the proto source code digest) and operation (use trials that match the proto operation name).

Use sourcecode or operation to use previous runs as input to the sequential optimization.

Regex matching for operation name filters

Use the characters ^ or $ at the start or end respectively of an operation filter to make the specified value a regular expression.

For example, this commands matches only run operations ending in train:

guild runs -Fo train$

Hydra integration

See Hydra examples for how Guild works with Hydra.

Multiple status chars with -S option

Use more than one status char with -S to expand the filter range. For example, to show only error AND terminated runs, use:

guild runs -Set

Platform independent PyPI distributions

Guild can now be installed on any architecture for Python 2.7 and 3.x.

Test coverage for Python 3.9

Guild is now tested on Python 3.9 on Windows, maxOS, and Linux.

Fixes

Guild not handling dict flag values in hparams TensorBoard support #276

windows-curses does not get installed alongside guildai on Windows #273

Config file bug. Single item list interpreted as the item itself. #271

Configuration files flag interface not working #263

Setting a flag for a required operation resource casts wrong if not alphanumeric #259

run state on remote not syncing to local #257

Lookup order for python_exe #250

Using global:<name> to import flag doesn’t work as described in the doc #248

Treat click as a normal dependency #247

2 Likes