Version 0.15.0
Major Features and Improvements
- Offered unified CLI for tfx pipeline actions on various orchestrators
including Apache Airflow, Apache Beam and Kubeflow. - Polished experimental interactive notebook execution and visualizations so
they are ready for use. - Added BulkInferrer component to TFX pipeline, and corresponding offline
inference taxi pipeline. - Introduced ImporterNode as a special TFX node to register external resource
into MLMD so that downstream nodes can use as input artifacts. An example
taxi_pipeline_importer.py
enabled by ImporterNode was added to showcase
the user journey of user-provided schema (issue #571). - Added experimental support for TFMA fairness indicator thresholds.
- Demonstrated DirectRunner multi-core processing in Chicago Taxi example,
including Airflow and Beam. - Introduced
PipelineConfig
andBaseComponentConfig
to control the
platform specific settings for pipelines and components. - Added a custom Executor of Pusher to push model to BigQuery ML for serving.
- Added KubernetesComponentLauncher to support launch ExecutorContainerSpec in a
Kubernetes cluster. - Made model validator executor forward compatible with TFMA change.
- Added Iris flowers classification example.
- Added support for serialization and deserialization of components.
- Made component launcher extensible to support launching components on
multiple platforms. - Simplified component package names.
- Introduced BaseNode as the base class of any node in a TFX pipeline DAG.
- Added docker component launcher to launch container component.
- Added support for specifying pipeline root in runtime when run on
KubeflowDagRunner. A default value can be provided when constructing the TFX
pipeline. - Added basic span support in ExampleGen to ingest file based data sources
that can be updated regularly by upstream. - Branched serving examples under chicago_taxi_pipeline/ from chicago_taxi/
example. - Supported beam arg 'direct_num_workers' for multi-processing on local.
- Improved naming of standard component inputs and outputs.
- Improved visualization functionality in the experimental TFX notebook
interface. - Allowed users to specify output file format when compiling TFX pipelines
using KubeflowDagRunner. - Introduced ResolverNode as a special TFX node to resolve input artifacts for
downstream nodes. ResolverNode is a convenient way to wrap TFX Resolver, a
logical unit for resolving input artifacts. - Added cifar-10 example to demonstrate image classification.
- Added container builder feature in the CLI tool for container-based custom
python components. This is specifically for the Kubeflow orchestration
engine, which requires containers built with the custom python code. - Demonstrated DirectRunner multi-core processing in Chicago Taxi example,
including Airflow and Beam. - Added Kubeflow artifact visualization of inputs, outputs and
execution properties for components using a Markdown file. Added Tensorboard
to Trainer components as well.
Bug fixes and other changes
- Bumped test dependency to kfp (Kubeflow Pipelines SDK) to be at version
0.1.31.2. - Fixed trainer executor to correctly make
transform_output
optional. - Updated Chicago Taxi example dependency tensorflow to version >=1.14.0.
- Updated Chicago Taxi example dependencies tensorflow-data-validation,
tensorflow-metadata, tensorflow-model-analysis, tensorflow-serving-api, and
tensorflow-transform to version >=0.14. - Updated Chicago Taxi example dependencies to Beam 2.14.0, Flink 1.8.1, Spark
2.4.3. - Adopted new recommended way to access component inputs/outputs as
component.outputs['output_name']
(previously, the syntax was
component.outputs.output_name
). - Updated Iris example to skip transform and use Keras model.
- Fixed the check for input artifact existence in base driver.
- Fixed bug in AI Platform Pusher that prevents pushes after first model, and
not being marked as default. - Replaced all usage of deprecated
tensorflow.logging
withabsl.logging
. - Used special user agent for all HTTP requests through googleapiclient and
apitools. - Transform component updated to use
tf.compat.v1
according to the TF 2.0
upgrading procedure. - TFX updated to use
tf.compat.v1
according to the TF 2.0 upgrading
procedure. - Added Kubeflow local example pipeline that executes components in-cluster.
- Fixed a bug that prevents updating execution type.
- Fixed a bug in model validator driver that reads across pipeline boundaries
when resolving latest blessed model. - Depended on
apache-beam[gcp]>=2.16,<3
- Depended on
ml-metadata>=0.15,<0.16
- Depended on
tensorflow>=1.15,<3
- Depended on
tensorflow-data-validation>=0.15,<0.16
- Depended on
tensorflow-model-analysis>=0.15.2,<0.16
- Depended on
tensorflow-transform>=0.15,<0.16
- Depended on 'tfx_bsl>=0.15.1,<0.16'
- Made launcher return execution information, containing populated inputs,
outputs, and execution id. - Updated the default configuration for accessing MLMD from pipelines running
in Kubeflow. - Updated Airflow developer tutorial
- CSVExampleGen: started using the CSV decoding utilities in
tfx-bsl
(tfx-bsl>=0.15.2
) - Added documentation for Fairness Indicators.
Deprecations
- Deprecated component_type in favor of type.
- Deprecated component_id in favor of id.
- Move beam_pipeline_args out of additional_pipeline_args as top level
pipeline param - Deprecated chicago_taxi folder, beam setup scripts and serving examples are
moved to chicago_taxi_pipeline folder.
Breaking changes
- Moved beam setup scripts from examples/chicago_taxi/ to
examples/chicago_taxi_pipeline/ - Moved interactive notebook classes into
tfx.orchestration.experimental
namespace. - Starting from 1.15, package
tensorflow
comes with GPU support. Users
won't need to choose betweentensorflow
andtensorflow-gpu
. If any GPU
devices are available, processes spawned by all TFX components will try to
utilize them; note that in rare cases, this may exhaust the memory of the
device(s). - Caveat:
tensorflow
2.0.0 is an exception and does not have GPU
support. Iftensorflow-gpu
2.0.0 is installed before installing
tfx
, it will be replaced withtensorflow
2.0.0.
Re-installtensorflow-gpu
2.0.0 if needed. - Caveat: MLMD schema auto-upgrade is now disabled by default. For users who
upgrades from 0.13 and do not want to lose the data in MLMD, please refer to
MLMD documentation
for guide to upgrade or downgrade MLMD database. Users who upgraded from TFX
0.14 should not be affected since there is not schema change between these
two versions.
For pipeline authors
- Deprecated the usage of
tf.contrib.training.HParams
in Trainer as it is
deprecated in TF 2.0. User module relying on member method of that class
will not be supported. Dot style property access will be the only supported
style from now on. - Any SavedModel produced by tf.Transform <=0.14 using any tf.contrib ops
(or tf.Transform ops that used tf.contrib ops such as tft.quantiles,
tft.bucketize, etc.) cannot be loaded with TF 2.0 since the contrib library
has been removed in 2.0. Please refer to this [issue]
(#838).
For component authors
Documentation updates
- Added conceptual info on Artifacts to guide/index.md