rszper commented on code in PR #25947:
URL: https://github.com/apache/beam/pull/25947#discussion_r1153591669


##########
website/www/site/content/en/documentation/ml/side-input-updates.md:
##########
@@ -0,0 +1,144 @@
+---
+title: "Auto Update ML models using WatchFilePattern"
+---
+<!--
+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.
+-->
+
+# Use WatchFilePattern to auto-update ML models in RunInference
+
+The pipeline in this example uses a 
[RunInference](https://beam.apache.org/documentation/transforms/python/elementwise/runinference/)
 `PTransform` to run inference on images using TensorFlow models. It uses a 
[side 
input](https://beam.apache.org/documentation/programming-guide/#side-inputs) 
`PCollection` that emits `ModelMetadata` to update the model.
+
+Using side inputs, you can update your model (which is passed in a 
`ModelHandler` configuration object) in real-time, even while the Beam pipeline 
is still running. This can be done either by leveraging one of Beam's provided 
patterns, such as the `WatchFilePattern`,
+or by configuring a custom side input `PCollection` that defines the logic for 
the model update.
+
+For more information about side inputs, see the [Side 
inputs](https://beam.apache.org/documentation/programming-guide/#side-inputs) 
section in the Apache Beam Programming Guide.
+
+This example uses 
[`WatchFilePattern`](https://beam.apache.org/releases/pydoc/current/apache_beam.ml.inference.utils.html#apache_beam.ml.inference.utils.WatchFilePattern)
 as a side input. `WatchFilePattern` is used to watch for the file updates 
matching the `file_pattern`
+based on timestamps. It emits the latest 
[`ModelMetadata`](https://beam.apache.org/documentation/transforms/python/elementwise/runinference/),
 which is used in
+the RunInference `PTransform` to automatically update the ML model without 
stopping the Beam pipeline.
+
+## Set up the source
+
+To read the image names, use a Pub/Sub topic as the source.
+ * The Pub/Sub topic emits a `UTF-8` encoded model path that is used to read 
and preprocess images to run the inference.
+
+## Models for image segmentation
+
+For the purpose of this example, use TensorFlow models saved in 
[HDF5](https://www.tensorflow.org/tutorials/keras/save_and_load#hdf5_format) 
format.
+
+
+## Pre-process images for inference
+The Pub/Sub topic emits an image path. We need to read and preprocess the 
image to use it for RunInference. The `read_image` function is used to read the 
image for inference.
+
+```python
+import io
+from PIL import Image
+from apache_beam.io.filesystems import FileSystems
+import numpy
+import tensorflow as tf
+
+def read_image(image_file_name):
+  with FileSystems().open(image_file_name, 'r') as file:
+    data = Image.open(io.BytesIO(file.read())).convert('RGB')
+  img = data.resize((224, 224))
+  img = numpy.array(img) / 255.0
+  img_tensor = tf.cast(tf.convert_to_tensor(img[...]), dtype=tf.float32)
+  return img_tensor
+```
+
+Now, let's jump into the pipeline code.
+
+**Pipeline steps**:
+1. Get the image names from the Pub/Sub topic.
+2. Read and pre-process the images using the `read_image` function.
+3. Pass the images to the RunInference `PTransform`. RunInference takes 
`model_handler` and `model_metadata_pcoll` as input parameters.
+
+For the 
[`model_handler`](https://github.com/apache/beam/blob/07f52a478174f8733c7efedb7189955142faa5fa/sdks/python/apache_beam/ml/inference/base.py#L308),
 we use 
[TFModelHandlerTensor](https://github.com/apache/beam/blob/186973b110d82838fb8e5ba27f0225a67c336591/sdks/python/apache_beam/ml/inference/tensorflow_inference.py#L184).
+```python
+from apache_beam.ml.inference.tensorflow_inference import TFModelHandlerTensor
+# initialize TFModelHandlerTensor with a .h5 model saved in a directory 
accessible by the pipeline.
+tf_model_handler = 
TFModelHandlerTensor(model_uri='gs://<your-bucket>/<model_path.h5>')
+```
+
+The `model_metadata_pcoll` is a [side 
input](https://beam.apache.org/documentation/programming-guide/#side-inputs) 
`PCollection` to the RunInference `PTransform`. This side input is used to 
update the models in the `model_handler` without needing to stop the beam 
pipeline.
+We will use `WatchFilePattern` as side input to watch a glob pattern matching 
`.h5` files.
+
+`model_metadata_pcoll` expects a `PCollection` of ModelMetadata compatible 
with 
[AsSingleton](https://beam.apache.org/releases/pydoc/2.4.0/apache_beam.pvalue.html#apache_beam.pvalue.AsSingleton)
 view. Because the pipeline uses `WatchFilePattern` as side input, it will take 
care of windowing and wrapping the output into `ModelMetadata`.

Review Comment:
   ```suggestion
   `model_metadata_pcoll` expects a `PCollection` of ModelMetadata compatible 
with the 
[AsSingleton](https://beam.apache.org/releases/pydoc/2.4.0/apache_beam.pvalue.html#apache_beam.pvalue.AsSingleton)
 class. Because the pipeline uses `WatchFilePattern` as side input, it will 
take care of windowing and wrapping the output into `ModelMetadata`.
   ```
   
   I'm not sure if this update is accurate, so please verify. I just find 
"compatible with AsSingleton view" confusing, so I'm trying to make what we 
mean here clearer.



##########
website/www/site/content/en/documentation/sdks/python-machine-learning.md:
##########
@@ -243,6 +243,23 @@ For more information, see the [`PredictionResult` 
documentation](https://github.
 For detailed instructions explaining how to build and run a Python pipeline 
that uses ML models, see the
 [Example RunInference API 
pipelines](https://github.com/apache/beam/tree/master/sdks/python/apache_beam/examples/inference)
 on GitHub.
 
+## Slowly-updating side input pattern to auto-update models used in 
RunInference
+To perform automatic updates of the models used with the RunInference 
`PTransform` without stopping the Beam pipeline, pass a 
[`ModelMetadata`](https://beam.apache.org/releases/pydoc/current/apache_beam.ml.inference.base.html#apache_beam.ml.inference.base.ModelMetadata)
 side input `PCollection` to the RunInference input parameter 
`model_metadata_pcoll`.
+
+`ModelMetdata` is a `NamedTuple` containing:
+  * `model_id`: Unique identifier for the model. This can be a file path or a 
URL where the model can be accessed. It is used to load the model for 
inference. The URL or file path must be in the compatible format so that the 
respective `ModelHandlers` can load the models without errors.
+
+    **For example**, `PyTorchModelHandler` initially loads a model using 
weights and a model class. If you pass in weights from a different model class 
when you update the model using side inputs, the model doesn't load properly, 
because it expects the weights from the original model class.
+  * `model_name`: Human-readable name for the model. You can use this name to 
identify the model in the metrics generated by the RunInference transform.
+
+Use cases:
+ * Use `WatchFilePattern` as side input to the RunInference `PTransform` to 
automatically update the ML model. For more information, see [Use 
`WatchFilePattern` as side input to auto-update ML models in 
RunInference](https://beam.apache.org/documentation/ml/side-input-updates).
+
+The side input `PCollection` must follow the 
[`AsSingleton`](https://beam.apache.org/releases/pydoc/current/apache_beam.pvalue.html?highlight=assingleton#apache_beam.pvalue.AsSingleton)
 view to avoid errors.
+
+**Note**: If the main `PCollection` emits inputs and a side input has yet to 
receive inputs, the main `PCollection` is buffered until there is
+            an update to the side input. This could happen with global 
windowed side inputs with data driven triggers, such as `AfterCount`, 
`AfterProcessingTime`. Until the side input is updated, emit the default or 
initial model ID that is used to pass the respective `ModelHandler` as a side 
input.

Review Comment:
   ```suggestion
               an update to the side input. This could happen with global 
windowed side inputs with data driven triggers, such as `AfterCount` and 
`AfterProcessingTime`. Until the side input is updated, emit the default or 
initial model ID that is used to pass the respective `ModelHandler` as a side 
input.
   ```



##########
website/www/site/content/en/documentation/ml/side-input-updates.md:
##########
@@ -0,0 +1,144 @@
+---
+title: "Auto Update ML models using WatchFilePattern"
+---
+<!--
+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.
+-->
+
+# Use WatchFilePattern to auto-update ML models in RunInference
+
+The pipeline in this example uses a 
[RunInference](https://beam.apache.org/documentation/transforms/python/elementwise/runinference/)
 `PTransform` to run inference on images using TensorFlow models. It uses a 
[side 
input](https://beam.apache.org/documentation/programming-guide/#side-inputs) 
`PCollection` that emits `ModelMetadata` to update the model.
+
+Using side inputs, you can update your model (which is passed in a 
`ModelHandler` configuration object) in real-time, even while the Beam pipeline 
is still running. This can be done either by leveraging one of Beam's provided 
patterns, such as the `WatchFilePattern`,
+or by configuring a custom side input `PCollection` that defines the logic for 
the model update.
+
+For more information about side inputs, see the [Side 
inputs](https://beam.apache.org/documentation/programming-guide/#side-inputs) 
section in the Apache Beam Programming Guide.
+
+This example uses 
[`WatchFilePattern`](https://beam.apache.org/releases/pydoc/current/apache_beam.ml.inference.utils.html#apache_beam.ml.inference.utils.WatchFilePattern)
 as a side input. `WatchFilePattern` is used to watch for the file updates 
matching the `file_pattern`
+based on timestamps. It emits the latest 
[`ModelMetadata`](https://beam.apache.org/documentation/transforms/python/elementwise/runinference/),
 which is used in
+the RunInference `PTransform` to automatically update the ML model without 
stopping the Beam pipeline.
+
+## Set up the source
+
+To read the image names, use a Pub/Sub topic as the source.
+ * The Pub/Sub topic emits a `UTF-8` encoded model path that is used to read 
and preprocess images to run the inference.

Review Comment:
   I don't think this needs to be a bullet point. These two sentences could 
just be one paragraph.



##########
website/www/site/content/en/documentation/sdks/python-machine-learning.md:
##########
@@ -243,6 +243,23 @@ For more information, see the [`PredictionResult` 
documentation](https://github.
 For detailed instructions explaining how to build and run a Python pipeline 
that uses ML models, see the
 [Example RunInference API 
pipelines](https://github.com/apache/beam/tree/master/sdks/python/apache_beam/examples/inference)
 on GitHub.
 
+## Slowly-updating side input pattern to auto-update models used in 
RunInference
+To perform automatic updates of the models used with the RunInference 
`PTransform` without stopping the Beam pipeline, pass a 
[`ModelMetadata`](https://beam.apache.org/releases/pydoc/current/apache_beam.ml.inference.base.html#apache_beam.ml.inference.base.ModelMetadata)
 side input `PCollection` to the RunInference input parameter 
`model_metadata_pcoll`.
+
+`ModelMetdata` is a `NamedTuple` containing:
+  * `model_id`: Unique identifier for the model. This can be a file path or a 
URL where the model can be accessed. It is used to load the model for 
inference. The URL or file path must be in the compatible format so that the 
respective `ModelHandlers` can load the models without errors.
+
+    **For example**, `PyTorchModelHandler` initially loads a model using 
weights and a model class. If you pass in weights from a different model class 
when you update the model using side inputs, the model doesn't load properly, 
because it expects the weights from the original model class.

Review Comment:
   ```suggestion
       For example, `PyTorchModelHandler` initially loads a model using weights 
and a model class. If you pass in weights from a different model class when you 
update the model using side inputs, the model doesn't load properly, because it 
expects the weights from the original model class.
   ```



-- 
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