standard_artifacts.py

# Copyright 2019 Google LLC. All Rights Reserved.
#
# 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.
"""A set of standard TFX Artifact types.

DO NOT USE THIS MODULE DIRECTLY. This module is a private module, and you should
use the redirected public module `tfx.v1.types.standard_artifacts`.

Note: the artifact definitions here are expected to change.
"""

import decimal
import math
from typing import Sequence

from absl import logging
from tfx.types import artifact
from tfx.types import standard_artifact_utils
from tfx.types import system_artifacts
from tfx.types import value_artifact
from tfx.utils import json_utils
from tfx.utils import pure_typing_utils

Artifact = artifact.Artifact
Property = artifact.Property
PropertyType = artifact.PropertyType
Dataset = system_artifacts.Dataset
SystemModel = system_artifacts.Model
Statistics = system_artifacts.Statistics
ValueArtifact = value_artifact.ValueArtifact

SPAN_PROPERTY = Property(type=PropertyType.INT)
VERSION_PROPERTY = Property(type=PropertyType.INT)
SPLIT_NAMES_PROPERTY = Property(type=PropertyType.STRING)
# (DEPRECATED. Do not use.) Value for a string-typed artifact.
STRING_VALUE_PROPERTY = Property(type=PropertyType.STRING)


class _TfxArtifact(Artifact):
  """TFX first-party component artifact definition.

  Do not construct directly, used for creating Channel, e.g.,
  ```
    Channel(type=standard_artifacts.Model)
  ```
  """

  def __init__(self, *args, **kwargs):
    """Construct TFX first-party component artifact."""
    # TODO(b/176795331): Refactor directory structure to make it clearer that
    # TFX-specific artifacts require the full "tfx" package be installed.
    #
    # Do not allow usage of TFX-specific artifact if only the core pipeline
    # SDK package is installed.
    try:
      import setuptools as _  # pytype: disable=import-error  # pylint: disable=g-import-not-at-top
      # Test import only when setuptools is available.
      try:
        # `extensions` is not included in ml_pipelines_sdk and doesn't have any
        # transitive import.
        import tfx.extensions as _  # type: ignore  # pylint: disable=g-import-not-at-top
      except ModuleNotFoundError as err:
        # The following condition detects exactly whether only the DSL package
        # is installed, and is bypassed when tests run in Bazel.
        raise RuntimeError('The "tfx" and all dependent packages need to be '
                           'installed to use this functionality.') from err
    except ModuleNotFoundError:
      pass

    super().__init__(*args, **kwargs)


class Examples(_TfxArtifact):
  """Artifact that contains the training data.

  Training data should be brought in to the TFX pipeline using components
  like ExampleGen. Data in Examples artifact is split and stored separately.
  The file and payload format must be specified as optional custom properties
  if not using default formats.
  Please see
  https://www.tensorflow.org/tfx/guide/examplegen#span_version_and_split to
  understand about span, version and splits.

  * Properties:
     - `span`: Integer to distinguish group of Examples.
     - `version`: Integer to represent updated data.
     - `splits`: A list of split names. For example, ["train", "test"].

  * File structure:
     - `{uri}/`
        - `Split-{split_name1}/`: Files for split
           - All direct children files are recognized as the data.
           - File format and payload format are determined by custom properties.
        - `Split-{split_name2}/`: Another split...

  * Commonly used custom properties of the Examples artifact:
    - `file_format`: a string that represents the file format. See
      tfx/components/util/tfxio_utils.py:make_tfxio for
      available values.
    - `payload_format`: int (enum) value of the data payload format.
      See tfx/proto/example_gen.proto:PayloadFormat for available formats.
  """
  TYPE_NAME = 'Examples'
  TYPE_ANNOTATION = Dataset
  PROPERTIES = {
      'span': SPAN_PROPERTY,
      'version': VERSION_PROPERTY,
      'split_names': SPLIT_NAMES_PROPERTY,
  }

  @property
  def splits(self) -> Sequence[str]:
    return standard_artifact_utils.decode_split_names(self.split_names)

  @splits.setter
  def splits(self, splits: Sequence[str]) -> None:
    if not pure_typing_utils.is_compatible(splits, Sequence[str]):
      raise TypeError(f'splits should be Sequence[str] but got {splits}')
    self.split_names = standard_artifact_utils.encode_split_names(list(splits))

  def path(self, *, split: str) -> str:
    """Path to the artifact URI's split subdirectory.

    This method DOES NOT create a directory path it returns; caller must make
    a directory of the returned path value before writing.

    Args:
      split: A name of the split, e.g. `"train"`, `"validation"`, `"test"`.

    Raises:
      ValueError: if the `split` is not in the `self.splits`.

    Returns:
      A path to `{self.uri}/Split-{split}`.
    """
    if split not in self.splits:
      raise ValueError(
          f'Split {split} not found in {self.splits=}. Did you forget to update'
          ' Examples.splits first?'
      )
    return standard_artifact_utils.get_split_uris([self], split)[0]


class ExampleAnomalies(_TfxArtifact):  # pylint: disable=missing-class-docstring
  TYPE_NAME = 'ExampleAnomalies'
  PROPERTIES = {
      'span': SPAN_PROPERTY,
      'split_names': SPLIT_NAMES_PROPERTY,
  }

  @property
  def splits(self) -> Sequence[str]:
    return standard_artifact_utils.decode_split_names(self.split_names)

  @splits.setter
  def splits(self, splits: Sequence[str]) -> None:
    if not pure_typing_utils.is_compatible(splits, Sequence[str]):
      raise TypeError(f'splits should be Sequence[str] but got {splits}')
    self.split_names = standard_artifact_utils.encode_split_names(list(splits))


class ExampleValidationMetrics(_TfxArtifact):  # pylint: disable=missing-class-docstring
  TYPE_NAME = 'ExampleValidationMetrics'
  PROPERTIES = {
      'span': SPAN_PROPERTY,
      'split_names': SPLIT_NAMES_PROPERTY,
  }

  @property
  def splits(self) -> Sequence[str]:
    return standard_artifact_utils.decode_split_names(self.split_names)

  @splits.setter
  def splits(self, splits: Sequence[str]) -> None:
    if not pure_typing_utils.is_compatible(splits, Sequence[str]):
      raise TypeError(f'splits should be Sequence[str] but got {splits}')
    self.split_names = standard_artifact_utils.encode_split_names(list(splits))


class ExampleStatistics(_TfxArtifact):  # pylint: disable=missing-class-docstring
  TYPE_NAME = 'ExampleStatistics'
  TYPE_ANNOTATION = Statistics
  PROPERTIES = {
      'span': SPAN_PROPERTY,
      'split_names': SPLIT_NAMES_PROPERTY,
  }

  @property
  def splits(self) -> Sequence[str]:
    return standard_artifact_utils.decode_split_names(self.split_names)

  @splits.setter
  def splits(self, splits: Sequence[str]) -> None:
    if not pure_typing_utils.is_compatible(splits, Sequence[str]):
      raise TypeError(f'splits should be Sequence[str] but got {splits}')
    self.split_names = standard_artifact_utils.encode_split_names(list(splits))


class ExamplesDiff(_TfxArtifact):
  TYPE_NAME = 'ExamplesDiff'


# TODO(b/158334890): deprecate ExternalArtifact.
class ExternalArtifact(_TfxArtifact):
  TYPE_NAME = 'ExternalArtifact'


class InferenceResult(_TfxArtifact):
  TYPE_NAME = 'InferenceResult'


class InfraBlessing(_TfxArtifact):
  TYPE_NAME = 'InfraBlessing'


class Model(_TfxArtifact):
  """Artifact that contains the actual persisted model.

  Training components stores the trained model like a saved model in this
  artifact. A `Model` artifact contains serialization of the trained model in
  one or more formats, each suitable for different usage (e.g. serving,
  evaluation), and serving environments.

  * File structure:
     - `{uri}/`
        - `Format-Serving/`: Model exported for serving.
           - `saved_model.pb`
           - Other actual model files.
        - `Format-TFMA/`: Model exported for evaluation.
           - `saved_model.pb`
           - Other actual model files.

  * Commonly used custom properties of the Model artifact:
  """
  TYPE_NAME = 'Model'
  TYPE_ANNOTATION = SystemModel


class ModelRun(_TfxArtifact):
  TYPE_NAME = 'ModelRun'


class ModelBlessing(_TfxArtifact):
  """Artifact that contains the evaluation of a trained model.

  This artifact is usually used with
  Conditional when determining
  whether to push this model on service or not.

  ```python
  # Run pusher if evaluator has blessed the model.
  with tfx.dsl.Cond(evaluator.outputs['blessing'].future()
                    [0].custom_property('blessed') == 1):
    pusher = Pusher(...)
  ```

  * File structure:
    - `{uri}/`
        - `BLESSED`: if the evaluator has blessed the model.
        - `NOT_BLESSED`: if the evaluator has not blessed the model.
    - See tfx/components/evaluator/executor.py for how to write
    ModelBlessing.

  * Commonly used custom properties of the ModelBlessing artifact:
    - `blessed`: int value that represents whether the evaluator has blessed its
    model or not.
  """
  TYPE_NAME = 'ModelBlessing'


class ModelEvaluation(_TfxArtifact):
  TYPE_NAME = 'ModelEvaluation'


class PushedModel(_TfxArtifact):
  TYPE_NAME = 'PushedModel'
  TYPE_ANNOTATION = SystemModel


class Schema(_TfxArtifact):
  """Artifact that contains the schema of the data.

  Schema artifact is used to store the
  schema of the data. The schema is a proto that describes the data, including
  the type of each feature, the range of values for each feature, and other
  properties. The schema is usually generated by the SchemaGen component, which
  uses the statistics of the data to infer the schema. The schema can be used by
  other components in the pipeline to validate the data and to generate models.

  * File structure:
    - `{uri}/`
      - `schema.pbtxt`: Text-proto format serialization of
        [tensorflow_metadata.proto.v0.schema.Schema](https://github.com/tensorflow/metadata/blob/master/tensorflow_metadata/proto/v0/schema.proto)
        proto message.
  """

  TYPE_NAME = 'Schema'


class TransformCache(_TfxArtifact):
  TYPE_NAME = 'TransformCache'


class JsonValue(ValueArtifact):
  """Artifacts representing a Jsonable value."""
  TYPE_NAME = 'JsonValue'

  def encode(self, value: json_utils.JsonableType) -> str:
    return json_utils.dumps(value)

  def decode(self, serialized_value: str) -> json_utils.JsonableType:
    return json_utils.loads(serialized_value)


class Bytes(ValueArtifact):
  """Artifacts representing raw bytes."""
  TYPE_NAME = 'Bytes'

  def encode(self, value: bytes):
    if not isinstance(value, bytes):
      raise TypeError('Expecting bytes but got value %s of type %s' %
                      (str(value), type(value)))
    return value

  def decode(self, serialized_value: bytes):
    return serialized_value


class String(ValueArtifact):
  """String-typed artifact.

  String value artifacts are encoded using UTF-8.
  """
  TYPE_NAME = 'String'

  # Note, currently we enforce unicode-encoded string.
  def encode(self, value: str) -> bytes:
    if not isinstance(value, str):
      raise TypeError('Expecting Text but got value %s of type %s' %
                      (str(value), type(value)))
    return value.encode('utf-8')

  def decode(self, serialized_value: bytes) -> str:
    return serialized_value.decode('utf-8')


class Boolean(ValueArtifact):
  """Artifacts representing a boolean.

  Boolean value artifacts are encoded as "1" for True and "0" for False.
  """
  TYPE_NAME = 'Boolean'

  def encode(self, value: bool):
    if not isinstance(value, bool):
      raise TypeError(
          f'Expecting bytes but got value {value} of type {type(value)}'
      )
    return b'1' if value else b'0'

  def decode(self, serialized_value: bytes):
    return int(serialized_value) != 0


class Integer(ValueArtifact):
  """Integer-typed artifact.

  Integer value artifacts are encoded as a decimal string.
  """
  TYPE_NAME = 'Integer'

  def encode(self, value: int) -> bytes:
    if not isinstance(value, int):
      raise TypeError(
          f'Expecting int but got value {value} of type {type(value)}'
      )
    return str(value).encode('utf-8')

  def decode(self, serialized_value: bytes) -> int:
    return int(serialized_value)


class Float(ValueArtifact):
  """Float-typed artifact.

  Float value artifacts are encoded using Python str() class. However,
  Nan and Infinity are handled separately. See string constants in the
  class.
  """
  TYPE_NAME = 'Float'

  _POSITIVE_INFINITY = float('Inf')
  _NEGATIVE_INFINITY = float('-Inf')

  _ENCODED_POSITIVE_INFINITY = 'Infinity'
  _ENCODED_NEGATIVE_INFINITY = '-Infinity'
  _ENCODED_NAN = 'NaN'

  def encode(self, value: float) -> bytes:
    if not isinstance(value, float):
      raise TypeError(
          f'Expecting float but got value {value} of type {type(value)}'
      )
    if math.isinf(value) or math.isnan(value):
      logging.warning(
          '! The number "%s" may be unsupported by non-python components.',
          value)
    str_value = str(value)
    # Special encoding for infinities and NaN to increase comatibility with
    # other languages.
    # Decoding works automatically.
    if math.isinf(value):
      if value >= 0:
        str_value = Float._ENCODED_POSITIVE_INFINITY
      else:
        str_value = Float._ENCODED_NEGATIVE_INFINITY
    if math.isnan(value):
      str_value = Float._ENCODED_NAN

    return str_value.encode('utf-8')

  def decode(self, serialized_value: bytes) -> float:
    result = float(serialized_value)

    # Check that the decoded value exactly matches the encoded string.
    # Note that float() can handle bytes, but Decimal() cannot.
    serialized_string = serialized_value.decode('utf-8')
    reserialized_string = str(result)
    is_exact = (decimal.Decimal(serialized_string) ==
                decimal.Decimal(reserialized_string))
    if not is_exact:
      logging.warning(
          'The number "%s" has lost precision when converted to float "%s"',
          serialized_value, reserialized_string)

    return result


class TransformGraph(_TfxArtifact):
  TYPE_NAME = 'TransformGraph'


class HyperParameters(_TfxArtifact):
  TYPE_NAME = 'HyperParameters'


class TunerResults(_TfxArtifact):
  TYPE_NAME = 'TunerResults'


# WIP and subject to change.
class DataView(_TfxArtifact):
  TYPE_NAME = 'DataView'


class Config(_TfxArtifact):
  TYPE_NAME = 'Config'