Tensorflow Model Analysis Architecture


The TensorFlow Model Analysis (TFMA) pipeline is depicted as follows:

TFMA Pipeline

The pipeline is made up of four main components:

  • Read Inputs
  • Extraction
  • Evaluation
  • Write Results

These components make use of two primary types: tfma.Extracts and tfma.evaluators.Evaluation. The type tfma.Extracts represents data that is extracted during pipeline processing whereas the type tfma.evaluators.Evaluation represents the output from evaluating the extracts at various points during the process of extraction. In order to provide a flexible API, these types are just dicts where the keys are defined (reserved for use) by different implementations. The types are defined as follows:

# Extracts represent data extracted during pipeline processing.
# For example, the PredictExtractor stores the data for the
# features, labels, and predictions under the keys "features",
# "labels", and "predictions".
Extracts = Dict[Text, Any]

# Evaluation represents the output from evaluating extracts at
# particular point in the pipeline. The evaluation outputs are
# keyed by their associated output type. For example, the serialized
# protos from evaluating metrics and plots will be stored under
# "metrics" and "plots" respectively.
Evaluation = Dict[Text, beam.pvalue.PCollection]

Note that tfma.Extracts are never written out directly they must always go through an evaluator to produce a tfma.evaluators.Evaluation that is then written out. Also note that tfma.Extracts are dicts that are stored in a beam.pvalue.PCollection (i.e. beam.PTransforms take as input beam.pvalue.PCollection[tfma.Extracts]) whereas a tfma.evaluators.Evaluation is a dict whose values are beam.pvalue.PCollections (i.e. beam.PTransforms take the dict itself as the argument for the beam.value.PCollection input). In other words the tfma.evaluators.Evaluation is used at pipeline construction time, but the tfma.Extracts are used at pipeline runtime.

Read Inputs

The ReadInputs stage is made up of a transform that takes raw inputs (tf.train.Example, CSV, ...) and converts them into extracts. Today the extracts are represented as raw input bytes stored under tfma.INPUT_KEY, however the extracts can be in any form that is compatible with the extraction pipeline -- meaning that it creates tfma.Extracts as output, and that those extracts are compatible with downstream extractors. It is up to the different extractors to clearly document what they require.


The extraction process is a list of beam.PTransforms that are run in series. The extractors take tfma.Extracts as input and return tfma.Extracts as output. The proto-typical extractor is the tfma.extractors.PredictExtractor which uses the input extract produced by the read inputs transform and runs it through a model to produce features, labels, and predictions extracts. Customized extractors can be inserted at any point provided their transforms conform to the tfma.Extracts in and tfma.Extracts out API. An extractor is defined as follows:

# An Extractor is a PTransform that takes Extracts as input and returns
# Extracts as output. A typical example is a PredictExtractor that receives
# an 'input' placeholder for input and adds additional 'features', 'labels',
# and 'predictions' extracts.
Extractor = NamedTuple('Extractor', [
    ('stage_name', Text),
    ('ptransform', beam.PTransform)])  # Extracts -> Extracts

Note that outside of very special cases, it is almost always the case that one tfma.Extracts in a beam.pvalue.PCollection will correspond to one example from the model.


Evaluation is the process of taking an extract and evaluating it. A common example is the tfma.evaluators.MetricsAndPlotsEvaluator which takes features, labels, and predictions as input, and then evaluates those inputs to produce metrics and plots data as output. While it is common to perform evaluation at the end of the extraction pipeline, there are use-cases that require evaluation earlier in the extraction process. As such evaluators are associated with the extractors whose output they should be evaluated against. An evaluator is defined as follows:

# An evaluator is a PTransform that takes Extracts as input and
# produces an Evaluation as output. A typical example of an evaluator
# is the MetricsAndPlotsEvaluator that takes the 'features', 'labels',
# and 'predictions' extracts from the PredictExtractor and evaluates
# them using post export metrics to produce serialized metrics and plots.
Evaluator = NamedTuple('Evaluator', [
  ('stage_name', Text),
  ('run_after', Text),              # Extractor.stage_name
  ('ptransform', beam.PTransform)]) # Extracts -> Evaluation

Notice that an evaluator is a beam.PTransform that takes tfma.Extracts as inputs. There is nothing stopping an implementation from performing additional transformations on the extracts as part of the evaluation process. This is exactly what the tfma.evaluators.MetricsAndPlotsEvaluator does. It takes the incoming features, labels, and predictions and runs them through tfma.slicer.FanoutSlices to group them by slices prior to performing the actual metrics and plots evaluation.

Also note that an evaluator can produce any output it wants. In the case of the tfma.evaluators.MetricsAndPlotsEvaluator the output is in the form of serialized protos.

Write Results

The WriteResults stage is where the evaluation output gets written out to disk. WriteResults uses writers to write out the data based on the output keys. For example, an tfma.evaluators.Evaluation may contain keys for 'metrics' and 'plots'. These would then be associated with the serialized protos called 'metrics' and 'plots'. The writers specify how to write out each file:

# A writer is a PTransform that takes evaluation output as input and
# serializes the associated PCollections of data to a sink.
Writer = NamedTuple('Writer', [
  ('stage_name', Text),
  ('ptransform', beam.PTransform)])    # Evaluation -> PDone

Since the tfma.evaluators.Evaluation passed to the writers contains the output for all of the evaluators combined, a tfma.writers.Write helper transform is provided that writers can use in their ptransform implementations to select the appropriate beam.PCollections based on an output key (see below for an example).


The tfma.run_model_analysis method takes extractors, evaluators, and writers arguments for customing the extractors, evaluators, and writers used by the pipeline. If no arguments are provided then tfma.default_extractors, tfma.default_evaluators, and tfma.default_writers are used by default.

Custom Extractors

To create a custom extractor, create a tfma.extractors.Extractor type that wraps a beam.PTransform taking tfma.Extracts as input and returning tfma.Extracts as output. Examples of extractors are available under tfma.extractors.

Custom Evaluators

To create a custom evaluator, create a tfma.evaluators.Evaluator type that wraps a beam.PTransform taking tfma.Extracts as input and returning tfma.evaluators.Evaluation as output. A very basic evaluator might just take the incoming tfma.Extracts and output them for storing in a table. This is exactly what the tfma.evaluators.AnalysisTableEvaluator does. A more complicated evaluator might perform additional processing and data aggregation. See the tfma.evaluators.MetricsAndPlotsEvaluator as an example..

Note that the tfma.evaluators.MetricsAndPlotsEvaluator itself can be customized to support custom metrics (see tfma.post_export_metrics for more details).

Custom Writers

To create a custom writer, create a tfma.writers.Writer type that wraps a beam.PTransform taking tfma.evaluators.Evaluation as input and returning beam.pvalue.PDone as output. The following is a basic example of a writer for writing out TFRecords containing metrics:

  stage_name='WriteTFRecord(%s)' % tfma.METRICS_KEY,

A writer's inputs depend on the output of the associated evaluator. For the above example, the output is a serialized proto produced by the tfma.evaluators.MetricsAndPlotsEvaluator. A writer for the tfma.evaluators.AnalysisTableEvaluator would be responsible for writing out a beam.pvalue.PCollection of tfma.Extracts.

Note that a writer is associated with the output of an evaluator via the output key used (e.g. tfma.METRICS_KEY, tfma.ANALYSIS_KEY, etc).

Step by Step Example

The following is an example of the steps involved in the extraction and evaluation pipeline when both the tfma.evaulators.MetricsAndPlotsEvaluator and tfma.evaluators.AnalysisTableEvaluator are used:



# Out
Extracts {
  'input': bytes                 # CSV, Proto, ...


  • tfma.extractors.PredictExtractor
# In:  ReadInputs Extracts
# Out:
Extracts {
  'input': bytes                    # CSV, Proto, ...
  'features': tensor_like           # Raw features
  'labels': tensor_like             # Labels
  'predictions': tensor_like        # Predictions
  • tfma.extractors.SliceKeyExtractor
# In: PredictExtractor Extracts
# Out:
Extracts {
  'features': tensor_like           # Raw features
  'labels': tensor_like             # Labels
  'predictions': tensor_like        # Predictions
  'slice_key': Tuple[bytes...]      # Slice
  • tfma.evaluators.MetricsAndPlotsEvaluator (run_after: SLICE_KEY_EXTRACTOR_STAGE_NAME)
# In: SliceKeyExtractor Extracts
# Out:
Evaluation {
  'metrics': PCollection[bytes]     # Serialized proto
  'plots': PCollection[bytes]       # Serialized proto
  • tfma.evaluators.AnalysisTableEvaluator (run_after: LAST_EXTRACTOR_STAGE_NAME)
# In: SliceKeyExtractor Extracts
# Out:
Evaluation {
  'analysis': PCollection[Extracts] # Final Extracts


# In:
Evaluation {
  'metrics': PCollection[bytes]     # Serialized proto
  'plots': PCollection[bytes]       # Serialized proto
  'analysis': PCollection[Extracts] # Final Extracts
# Out: metrics, plots, and analysis files