[0.5.0] - 2021-02-21
Added
- A new
long_parameters_strategy
key is added in themlflow.yml
(under in the hook/node section). You can specify different strategies (fail
,truncate
ortag
) to handle parameters over 250 characters which cause crashes for some mlflow backend. (#69) - Add an
env
parameter tokedro mlflow init
command to specify under whichconf/
subfolder themlflow.yml
should be created. (#159) - The input parameters of the
inference
pipeline of aPipelineML
object are now automatically pickle-ised and converted as artifacts. (#158) - Detailed documentation on how to use
pipeline_ml_factory
function, and more generally on how to usekedro-mlflow
as mlops framework. This comes from an example repokedro-mlflow-tutorial
. (#16)
Fixed
- Pin the kedro version to force it to be strictly inferior to
0.17
which is not compatible with currentkedro-mlflow
version (#143) - It is no longer assumed for the project to run that the
mlflow.yml
is located underconf/base
. The project will run as soon as the configuration file is discovered by the registered ConfigLoader (#159)
Changed
- The
KedroPipelineModel.load_context()
method now loads all theDataSets
in memory in theDataCatalog
. It is also now possible to specify therunner
to execute the model as well as thecopy_mode
when executing the inference pipeline (instead of deepcopying the datasets between each nodes which is kedro's default). This makes the API serving withmlflow serve
command considerably faster (~20 times faster) for models which need compiling (e.g. keras, tensorflow ...) (#133) - The CLI projects commands are now always accessible even if you have not called
kedro mlflow init
yet to create amlflow.yml
configuration file (#159)
�