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]
