rszper commented on code in PR #22250: URL: https://github.com/apache/beam/pull/22250#discussion_r920291223
########## website/www/site/content/en/documentation/sdks/python-machine-learning.md: ########## @@ -0,0 +1,186 @@ +--- +type: languages +title: "Apache Beam Python Machine Learning" +--- +<!-- +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +--> + +# Machine Learning + +You can use Apache Beam with the RunInference API to use machine learning (ML) models to do local and remote inference with batch and streaming pipelines. Starting with Apache Beam 2.40.0, PyTorch and Scikit-learn frameworks are supported. You can create multiple types of transforms using the RunInference API: the API takes multiple types of setup parameters from model handlers, and the parameter type determines the model implementation. + +## Why use the RunInference API? + +RunInference leverages existing Apache Beam concepts, such as the the `BatchElements` transform and the `Shared` class, and it allows you to build multi-model pipelines. In addition, the RunInference API has built in capabilities for dealing with [keyed values](#use-the-prediction-results-object). + +### BatchElements PTransform + +To take advantage of the optimizations of vectorized inference that many models implement, we added the `BatchElements` transform as an intermediate step before making the prediction for the model. This transform batches elements together. The resulting batch is used to make the appropriate transformation for the particular framework of RunInference. For example, for numpy `ndarrays`, we call `numpy.stack()`, and for torch `Tensor` elements, we call `torch.stack()`. + +To customize the settings for `beam.BatchElements`, in `ModelHandler`, override the `batch_elements_kwargs` function. For example, use `min_batch_size` to set the lowest number of elements per batch or `max_batch_size` to set the highest number of elements per batch. + +For more information, see the [`BatchElements` transform documentation](https://beam.apache.org/releases/pydoc/current/apache_beam.transforms.util.html#apache_beam.transforms.util.BatchElements). + +### Shared helper class + +Instead of loading a model for each thread in the process, we use the `Shared` class, which allows us to load one model that is shared across all threads of each worker in a DoFn. For more information, see the +[`Shared` class documentation](https://github.com/apache/beam/blob/master/sdks/python/apache_beam/utils/shared.py#L20). + +### Multi-model pipelines + +The RunInference API can be composed into multi-model pipelines. Multi-model pipelines are useful for A/B testing and for building out ensembles for tokenization, sentence segmentation, part-of-speech tagging, named entity extraction, language detection, coreference resolution, and more. Review Comment: I'm agnostic about whether to have specifics or a generalization. If we make it more generic, what would you suggest? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
