About Beam ML
| Javadoc |
You can use Apache Beam to:
- Process large volumes of data, both for preprocessing and for inference.
- Experiment with your data during the exploration phase of your project.
- Upscale your data pipelines as part of your ML ops ecosystem in a production environment.
- Run your model in production on a varying data load, both in batch and streaming.
AI/ML workloads
You can use Apache Beam for data validation, data preprocessing, model validation, and model deployment and inference.
- Data ingestion: Incoming new data is either stored in your file system or database, or published to a messaging queue.
- Data validation: After you receieve your data, check the quality of the data. For example, you might want to detect outliers and calculate standard deviations and class distributions.
- Data preprocessing: After you validate your data, transform the data so that it’s ready to use to train your model.
- Model training: When your data is ready, train your AI/ML model. This step is typically repeated multiple times, depending on the quality of your trained model.
- Model validation: Before you deploy your model, validate its performance and accuracy.
- Model deployment: Deploy your model, using it to run inference on new or existing data.
To keep your model up to date and performing well as your data grows and evolves, run these steps multiple times. In addition, you can apply ML ops to your project to automate the AI/ML workflows throughout the model and data lifecycle. Use orchestrators to automate this flow and to handle the transition between the different building blocks in your project.
Use RunInference
The RunInference API is a PTransform
optimized for machine learning inferences that lets you efficiently use ML models in your pipelines. The API includes the following features:
- To efficiently feed your model, dynamically batches inputs based on pipeline throughput using Apache Beam’s
BatchElements
transform. - To balance memory and throughput usage, determines the optimal number of models to load using a central model manager. Shares these models across threads and processes as needed to maximize throughput.
- Ensures that your pipeline uses the most recently deployed version of your model with the Automatic model refresh feature.
- Supports multiple frameworks and model hubs, including Tensorflow, Pytorch, Sklearn, XGBoost, Hugging Face, TensorFlow Hub, Vertex AI, TensorRT, and ONNX.
- Supports arbitrary frameworks using a custom model handler.
- Supports multi-model pipelines.
- Lets you use GPUs on supported runners to increase inference speed. For more information, see GPUs with Dataflow in the Dataflow documentation.
Support and limitations
- The RunInference API is supported in Apache Beam 2.40.0 and later versions.
- Model handlers are available for PyTorch, scikit-learn, TensorFlow, Hugging Face, Vertex AI, ONNX, TensorRT, and XGBoost. You can also use a custom model handler.
- The RunInference API supports batch and streaming pipelines.
- The RunInference API supports both remote inference and inteference local to the runner worker.
BatchElements PTransform
To take advantage of the optimizations of vectorized inference that many models implement, the BatchElements
transform is used 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.
Shared helper class
Using the Shared
class within the RunInference implementation makes it possible to load the model only once per process and share it with all DoFn instances created in that process. This feature reduces memory consumption and model loading time. For more information, see the
Shared
class documentation.
Modify a Python pipeline to use an ML model
To use the RunInference transform, add the following code to your pipeline:
from apache_beam.ml.inference.base import RunInference
with pipeline as p:
predictions = ( p | 'Read' >> beam.ReadFromSource('a_source')
| 'RunInference' >> RunInference(<model_handler>)
Replace model_handler
with the model handler setup code.
To import models, you need to configure a ModelHandler
object that wraps the underlying model. Which model handler you import depends on the framework and type of data structure that contains the inputs. The ModelHandler
object also allows you to set environment variables needed for inference using the env_vars
keyword argument. The following examples show some model handlers that you might want to import.
from apache_beam.ml.inference.sklearn_inference import SklearnModelHandlerNumpy
from apache_beam.ml.inference.sklearn_inference import SklearnModelHandlerPandas
from apache_beam.ml.inference.pytorch_inference import PytorchModelHandlerTensor
from apache_beam.ml.inference.pytorch_inference import PytorchModelHandlerKeyedTensor
from tfx_bsl.public.beam.run_inference import CreateModelHandler
Use pre-trained models
The section provides requirements for using pre-trained models with PyTorch, Scikit-learn, and Tensorflow.
PyTorch
You need to provide a path to a file that contains the model’s saved weights. This path must be accessible by the pipeline. To use pre-trained models with the RunInference API and the PyTorch framework, complete the following steps:
- Download the pre-trained weights and host them in a location that the pipeline can access.
- Pass the path of the model weights to the PyTorch
ModelHandler
by using the following code:state_dict_path=<path_to_weights>
.
See this notebook that illustrates running PyTorch models with Apache Beam.
Scikit-learn
You need to provide a path to a file that contains the pickled Scikit-learn model. This path must be accessible by the pipeline. To use pre-trained models with the RunInference API and the Scikit-learn framework, complete the following steps:
- Download the pickled model class and host it in a location that the pipeline can access.
- Pass the path of the model to the Sklearn
ModelHandler
by using the following code:model_uri=<path_to_pickled_file>
andmodel_file_type: <ModelFileType>
, where you can specifyModelFileType.PICKLE
orModelFileType.JOBLIB
, depending on how the model was serialized.
See this notebook that illustrates running Scikit-learn models with Apache Beam.
TensorFlow
To use TensorFlow with the RunInference API, you have two options:
- Use the built-in TensorFlow Model Handlers in Apache Beam SDK -
TFModelHandlerNumpy
andTFModelHandlerTensor
.- Depending on the type of input for your model, use
TFModelHandlerNumpy
fornumpy
input andTFModelHandlerTensor
fortf.Tensor
input respectively. - Use tensorflow 2.7 or later.
- Pass the path of the model to the TensorFlow
ModelHandler
by usingmodel_uri=<path_to_trained_model>
. - Alternatively, you can pass the path to saved weights of the trained model, a function to build the model using
create_model_fn=<function>
, and set themodel_type=ModelType.SAVED_WEIGHTS
. See this notebook that illustrates running Tensorflow models with Built-in model handlers.
- Depending on the type of input for your model, use
- Using
tfx_bsl
.- Use this approach if your model input is of type
tf.Example
. - Use
tfx_bsl
version 1.10.0 or later. - Create a model handler using
tfx_bsl.public.beam.run_inference.CreateModelHandler()
. - Use the model handler with the
apache_beam.ml.inference.base.RunInference
transform. See this notebook that illustrates running TensorFlow models with Apache Beam and tfx-bsl.
- Use this approach if your model input is of type
Use custom models
If you would like to use a model that isn’t specified by one of the supported frameworks, the RunInference API is designed flexibly to allow you to use any custom machine learning models.
You only need to create your own ModelHandler
or KeyedModelHandler
with logic to load your model and use it to run the inference.
A simple example can be found in this notebook.
The load_model
method shows how to load the model using a popular spaCy
package while run_inference
shows how to run the inference on a batch of examples.
RunInference Patterns
This section suggests patterns and best practices that you can use to make your inference pipelines simpler, more robust, and more efficient.
Use a keyed ModelHandler object
If a key is attached to the examples, wrap KeyedModelHandler
around the ModelHandler
object:
from apache_beam.ml.inference.base import KeyedModelHandler
keyed_model_handler = KeyedModelHandler(PytorchModelHandlerTensor(...))
with pipeline as p:
data = p | beam.Create([
('img1', torch.tensor([[1,2,3],[4,5,6],...])),
('img2', torch.tensor([[1,2,3],[4,5,6],...])),
('img3', torch.tensor([[1,2,3],[4,5,6],...])),
])
predictions = data | RunInference(keyed_model_handler)
If you are unsure if your data is keyed, you can use MaybeKeyedModelHandler
.
You can also use a KeyedModelHandler
to load several different models based on their associated key.
The following example loads a model by using config1
. That model is used for inference for all examples associated
with key1
. It loads a second model by using config2
. That model is used for all examples associated with key2
and key3
.
from apache_beam.ml.inference.base import KeyedModelHandler
keyed_model_handler = KeyedModelHandler([
KeyModelMapping(['key1'], PytorchModelHandlerTensor(<config1>)),
KeyModelMapping(['key2', 'key3'], PytorchModelHandlerTensor(<config2>))
])
with pipeline as p:
data = p | beam.Create([
('key1', torch.tensor([[1,2,3],[4,5,6],...])),
('key2', torch.tensor([[1,2,3],[4,5,6],...])),
('key3', torch.tensor([[1,2,3],[4,5,6],...])),
])
predictions = data | RunInference(keyed_model_handler)
For a more detailed example, see the notebook Run ML inference with multiple differently-trained models.
Loading multiple models at the same times increases the risk of out of memory errors (OOMs). By default, KeyedModelHandler
doesn’t
limit the number of models loaded into memory at the same time. If the models don’t all fit into memory,
your pipeline might fail with an out of memory error. To avoid this issue, use the max_models_per_worker_hint
parameter
to set the maximum number of models that can be loaded into memory at the same time.
The following example loads at most two models per SDK worker process at a time. It unloads models that aren’t currently in use.
mhs = [
KeyModelMapping(['key1'], PytorchModelHandlerTensor(<config1>)),
KeyModelMapping(['key2', 'key3'], PytorchModelHandlerTensor(<config2>)),
KeyModelMapping(['key4'], PytorchModelHandlerTensor(<config3>)),
KeyModelMapping(['key5', 'key6', 'key7'], PytorchModelHandlerTensor(<config4>)),
]
keyed_model_handler = KeyedModelHandler(mhs, max_models_per_worker_hint=2)
Runners that have multiple SDK worker processes on a given machine load at most
max_models_per_worker_hint*<num worker processes>
models onto the machine.
Leave enough space for the models and any additional memory needs from other transforms. Because the memory might not be released immediately after a model is offloaded, leaving an additional buffer is recommended.
Note: Having many models but a small max_models_per_worker_hint
can cause memory thrashing, where
a large amount of execution time is used to swap models in and out of memory. To reduce the likelihood and impact
of memory thrashing, if you’re using a distributed runner, insert a
GroupByKey
transform before your
inference step. The GroupByKey
transform reduces thrashing by ensuring that elements with the same key and model are
collocated on the same worker.
For more information, see KeyedModelHander
.
Use the PredictionResult object
When doing a prediction in Apache Beam, the output PCollection
includes both the keys of the input examples and the inferences. Including both these items in the output allows you to find the input that determined the predictions.
The PredictionResult
object is a NamedTuple
that contains both the input and the inferences, named example
and inference
, respectively. When keys are passed with the input data to the RunInference transform, the output PCollection
returns a Tuple[str, PredictionResult]
, which is the key and the PredictionResult
object. Your pipeline interacts with a PredictionResult
object in steps after the RunInference transform.
class PostProcessor(beam.DoFn):
def process(self, element: Tuple[str, PredictionResult]):
key, prediction_result = element
inputs = prediction_result.example
predictions = prediction_result.inference
# Post-processing logic
result = ...
yield (key, result)
with pipeline as p:
output = (
p | 'Read' >> beam.ReadFromSource('a_source')
| 'PyTorchRunInference' >> RunInference(<keyed_model_handler>)
| 'ProcessOutput' >> beam.ParDo(PostProcessor()))
If you need to use this object explicitly, include the following line in your pipeline to import the object:
from apache_beam.ml.inference.base import PredictionResult
For more information, see the PredictionResult
documentation.
Automatic model refresh
To automatically update the model being used with the RunInference PTransform
without stopping the pipeline, pass a 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 respectiveModelHandlers
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 RunInferencePTransform
to automatically update the ML model. For more information, see UseWatchFilePattern
as side input to auto-update ML models in RunInference.
The side input PCollection
must follow the 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
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.
Preprocess and postprocess your records
With RunInference, you can add preprocessing and postprocessing operations to your transform.
To apply preprocessing operations, use with_preprocess_fn
on your model handler:
inference = pcoll | RunInference(model_handler.with_preprocess_fn(lambda x : do_something(x)))
To apply postprocessing operations, use with_postprocess_fn
on your model handler:
inference = pcoll | RunInference(model_handler.with_postprocess_fn(lambda x : do_something_to_result(x)))
You can also chain multiple pre- and postprocessing operations:
inference = pcoll | RunInference(
model_handler.with_preprocess_fn(
lambda x : do_something(x)
).with_preprocess_fn(
lambda x : do_something_else(x)
).with_postprocess_fn(
lambda x : do_something_after_inference(x)
).with_postprocess_fn(
lambda x : do_something_else_after_inference(x)
))
The preprocessing function is run before batching and inference. This function maps your input PCollection
to the base input type of the model handler. If you apply multiple preprocessing functions, they run on your original
PCollection
in the order of last applied to first applied.
The postprocessing function runs after inference. This function maps the output type of the base model handler to your desired output type. If you apply multiple postprocessing functions, they run on your original inference result in the order of first applied to last applied.
Handle errors
To handle errors robustly while using RunInference, you can use a dead-letter queue. The dead-letter queue outputs failed records into a separate PCollection
for further processing.
This PCollection
can then be analyzed and sent to a storage system, where it can be reviewed and resubmitted to the pipeline, or discarded.
RunInference has built-in support for dead-letter queues. You can use a dead-letter queue by applying with_exception_handling
to your RunInference transform:
main, other = pcoll | RunInference(model_handler).with_exception_handling()
other.failed_inferences | beam.Map(print) # insert logic to handle failed records here
You can also apply this pattern to RunInference transforms with associated pre- and postprocessing operations:
main, other = pcoll | RunInference(model_handler.with_preprocess_fn(f1).with_postprocess_fn(f2)).with_exception_handling()
other.failed_preprocessing[0] | beam.Map(print) # handles failed preprocess operations, indexed in the order in which they were applied
other.failed_inferences | beam.Map(print) # handles failed inferences
other.failed_postprocessing[0] | beam.Map(print) # handles failed postprocess operations, indexed in the order in which they were applied
Run inference from a Java pipeline
The RunInference API is available with the Beam Java SDK versions 2.41.0 and later through Apache Beam’s Multi-language Pipelines framework. For information about the Java wrapper transform, see RunInference.java. To try it out, see the Java Sklearn Mnist Classification example. Additionally, see Using RunInference from Java SDK for an example of a composite Python transform that uses the RunInference API along with preprocessing and postprocessing from a Beam Java SDK pipeline.
Custom Inference
The RunInference API doesn’t currently support making remote inference calls using, for example, the Natural Language API or the Cloud Vision API. Therefore, in order to use these remote APIs with Apache Beam, you need to write custom inference calls. The Remote inference in Apache Beam notebook shows how to implement a custom remote inference call using beam.DoFn
. When you implement a remote inference for real life projects, consider the following factors:
API quotas and the heavy load you might incur on your external API. To optimize the calls to an external API, you can confgure
PipelineOptions
to limit the parallel calls to the external remote API.Be prepared to encounter, identify, and handle failure as gracefully as possible. Use techniques like exponential backoff and dead-letter queues (unprocessed messages queues).
When running inference with an external API, batch your input together to allow for more efficient execution.
Consider monitoring and measuring the performance of a pipeline when deploying, because monitoring can provide insight into the status and health of the application.
Multi-model pipelines
Use the RunInference transform to add multiple inference models to your pipeline. Multi-model pipelines can be useful for A/B testing or for building out cascade models made up of models that perform tokenization, sentence segmentation, part-of-speech tagging, named entity extraction, language detection, coreference resolution, and more. For more information, see Multi-model pipelines.
A/B Pattern
with pipeline as p:
data = p | 'Read' >> beam.ReadFromSource('a_source')
model_a_predictions = data | RunInference(<model_handler_A>)
model_b_predictions = data | RunInference(<model_handler_B>)
Where model_handler_A
and model_handler_B
are the model handler setup code.
Cascade Pattern
with pipeline as p:
data = p | 'Read' >> beam.ReadFromSource('a_source')
model_a_predictions = data | RunInference(<model_handler_A>)
model_b_predictions = model_a_predictions | beam.Map(some_post_processing) | RunInference(<model_handler_B>)
Where model_handler_A
and model_handler_B
are the model handler setup code.
Use Resource Hints for Different Model Requirements
When using multiple models in a single pipeline, different models may have different memory or worker SKU requirements. Resource hints allow you to provide information to a runner about the compute resource requirements for each step in your pipeline.
For example, the following snippet extends the previous cascade pattern with hints for each RunInference call to specify RAM and hardware accelerator requirements:
with pipeline as p:
data = p | 'Read' >> beam.ReadFromSource('a_source')
model_a_predictions = data | RunInference(<model_handler_A>).with_resource_hints(min_ram="20GB")
model_b_predictions = model_a_predictions
| beam.Map(some_post_processing)
| RunInference(<model_handler_B>).with_resource_hints(
min_ram="4GB",
accelerator="type:nvidia-tesla-k80;count:1;install-nvidia-driver")
For more information on resource hints, see Resource hints.
Model validation
Model validation allows you to benchmark your model’s performance against a previously unseen dataset. You can extract chosen metrics, create visualizations, log metadata, and compare the performance of different models with the end goal of validating whether your model is ready to deploy. Beam provides support for running model evaluation on a TensorFlow model directly inside your pipeline.
The ML model evaluation page shows how to integrate model evaluation as part of your pipeline by using TensorFlow Model Analysis (TFMA).
Troubleshooting
If you run into problems with your pipeline or job, this section lists issues that you might encounter and provides suggestions for how to fix them.
Unable to batch tensor elements
RunInference uses dynamic batching. However, the RunInference API cannot batch tensor elements of different sizes, so samples passed to the RunInference
transform must be the same dimension or length. If you provide images of different sizes or word embeddings of different lengths, the following error might occur:
File "/beam/sdks/python/apache_beam/ml/inference/pytorch_inference.py", line 232, in run_inference batched_tensors = torch.stack(key_to_tensor_list[key]) RuntimeError: stack expects each tensor to be equal size, but got [12] at entry 0 and [10] at entry 1 [while running 'PyTorchRunInference/ParDo(_RunInferenceDoFn)']
To avoid this issue, either use elements of the same size, or disable batching.
Option 1: Use elements of the same size
Use elements of the same size or resize the inputs. For computer vision applications, resize image inputs so that they have the same dimensions. For natural language processing (NLP) applications that have text of varying length, resize the text or word embeddings to make them the same length. When working with texts of varying length, resizing might not be possible. In this scenario, you could disable batching (see option 2).
Option 2: Disable batching
Disable batching by overriding the batch_elements_kwargs
function in your ModelHandler and setting the maximum batch size (max_batch_size
) to one: max_batch_size=1
. For more information, see
BatchElements PTransforms. For an example, see our language modeling example.
Related links
- RunInference transforms
- RunInference API pipeline examples
- RunInference public codelab
- RunInference notebooks
- RunInference benchmarks
| Javadoc |
Last updated on 2024/11/14
Have you found everything you were looking for?
Was it all useful and clear? Is there anything that you would like to change? Let us know!