tvalentyn commented on code in PR #22250:
URL: https://github.com/apache/beam/pull/22250#discussion_r922486663


##########
website/www/site/content/en/documentation/sdks/python-machine-learning.md:
##########
@@ -0,0 +1,201 @@
+---
+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 takes advantage of existing Apache Beam concepts, such as the the 
`BatchElements` transform and the `Shared` class, to enable you to use models 
in your pipelines to create transforms optimized for machine learning 
inferences. The ability to create arbitrarily complex workflow graphs also 
allows you to build multi-model pipelines.
+
+### 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 batched elements are then applied with a 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

Review Comment:
   How about: 
   
   Using the `Shared` class within RunInference implementation allows us to 
load the model  only once per process and share it with all DoFn instances 
created in that process. This reduces the memory consumption and model loading 
time. 



-- 
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]

Reply via email to