saved_model_metadata_builder.py

# Copyright 2021 Google LLC
#
# 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.

"""Metadata builder for models built with Tensorflow 2.X.

This metadata builder supports all three flavors of Keras model interface:
sequential, functional, and subclassed.

Inputs and outputs of a model can be inferred from a saved model. Using the
provided saved model signatures, we create input and output metadata. Users have
the option to remove and modify these metadata. Users have the option to resave
their model with the metadata file or get the metadata from the builder to save
it themselves.

This builder infers metadata for all inputs and outputs. However, explainability
service supports only single output. Users need to use remove_output_metadata
and get_metadata functions to remove the ones they don't need.
"""
from typing import Optional, List, Dict, Union, Any, Tuple
import tensorflow as tf
from explainable_ai_sdk.common import explain_metadata
from explainable_ai_sdk.common import types
from explainable_ai_sdk.metadata import constants
from explainable_ai_sdk.metadata import metadata_builder
from explainable_ai_sdk.metadata import parameters
from explainable_ai_sdk.metadata import utils


class SavedModelMetadataBuilder(metadata_builder.MetadataBuilder):
  """Class for generating metadata for a model built with TF 2.X Keras API."""

  def __init__(
      self,
      model_path: str,
      signature_name: str = tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY,
      outputs_to_explain: Optional[List[str]] = (),
      **kwargs) -> None:  # pytype: disable=annotation-type-mismatch
    """Initializes a SavedModelMetadataBuilder object.

    Args:
      model_path: Path to load the saved model from.
      signature_name: Name of the signature to be explained. Inputs and outputs
        of this signature will be written in the metadata. If not provided, the
        default signature will be used.
      outputs_to_explain: List of output names to explain. Only single output is
        supported for now. Hence, the list should contain one element. This
        parameter is required if the model signature (provided via
        signature_name) specifies multiple outputs.
      **kwargs: Any keyword arguments to be passed to tf.saved_model.save()
        function.
    """
    if outputs_to_explain and len(outputs_to_explain) > 1:
      raise ValueError('"outputs_to_explain" can only contain 1 element.\n'
                       'Got: %s' % len(outputs_to_explain))
    self._explain_output = outputs_to_explain
    self._saved_model_args = kwargs
    self._loaded_model = tf.saved_model.load(model_path)
    self._inputs, self._outputs = self._infer_metadata_entries_from_model(
        signature_name)

  def _infer_metadata_entries_from_model(
      self, signature_name: str
  ) -> Tuple[Dict[str, explain_metadata.InputMetadata],
             Dict[str, explain_metadata.OutputMetadata]]:
    """Infers metadata inputs and outputs."""
    loaded_sig = self._loaded_model.signatures[signature_name]
    _, input_sig = loaded_sig.structured_input_signature
    output_sig = loaded_sig.structured_outputs
    input_mds = {}
    for name, tensor_spec in input_sig.items():
      if tensor_spec.dtype.is_floating:
        input_mds[name] = explain_metadata.InputMetadata(name, name)
      else:
        input_mds[name] = explain_metadata.InputMetadata(
            name, name, modality=explain_metadata.Modality.CATEGORICAL)

    if not self._explain_output and len(output_sig) > 1:
      raise ValueError("Signature has multiple outputs. You must specify which"
                       " output to explain via 'outputs_to_explain' parameter.")
    for name in output_sig:
      if not self._explain_output or self._explain_output[0] == name:
        output_mds = {name: explain_metadata.OutputMetadata(name, name)}
        break
    else:
      raise ValueError("Specified output name cannot be found in given"
                       " signature outputs.")
    return input_mds, output_mds

  def set_numeric_metadata(
      self,
      input_name: str,
      new_name: Optional[str] = None,
      input_baselines: Optional[List[Union[types.TensorValue,
                                           types.Tensor]]] = None,
      index_feature_mapping: Optional[List[str]] = None) -> None:
    """Sets an existing metadata identified by input as numeric with params.

    Args:
      input_name: Input name in the metadata to be set as numeric.
      new_name: Optional (unique) new name for this feature.
      input_baselines: A list of baseline values. Each baseline value can be a
        single entity or of the same shape as the model_input (except for the
        batch dimension).
      index_feature_mapping: A list of feature names for each index in the input
        tensor.

    Raises:
      ValueError: If input_name cannot be found in the metadata.
    """
    if input_name not in self._inputs:
      raise ValueError("Input with with name '%s' does not exist." % input_name)
    name = new_name if new_name else input_name
    tensor_name = self._inputs.pop(input_name).input_tensor_name

    if index_feature_mapping:
      encoding = explain_metadata.Encoding.BAG_OF_FEATURES
    else:
      encoding = explain_metadata.Encoding.IDENTITY
    self._inputs[name] = explain_metadata.InputMetadata(
        name=name,
        input_tensor_name=tensor_name,
        input_baselines=input_baselines,
        index_feature_mapping=index_feature_mapping,
        encoding=encoding)

  def set_categorical_metadata(
      self,
      input_name: str,
      new_name: Optional[str] = None,
      encoded_name: Optional[str] = None,
      encoding: str = explain_metadata.Encoding.IDENTITY,
      input_baselines: Optional[List[Union[types.TensorValue,
                                           types.Tensor]]] = None,
      encoded_baselines: Optional[List[Union[types.TensorValue,
                                             types.Tensor]]] = None) -> None:
    """Sets an existing metadata identified by input as categorical with params.

    Args:
      input_name: Input name in the metadata to be set as numeric.
      new_name: Optional (unique) new name for this feature.
      encoded_name: Optional name of the tensor, which is the encoded version of
        the input tensor. It is potentially an output of an encoding function
        such as embedding. Each encoded_name should map to a unique input.
      encoding: Encoding type if encoded_tensor is provided. Possible values are
        {identity, bag_of_features, bag_of_features_sparse, indicator,
        combined_embedding, concat_embedding}.
      input_baselines: A list of baseline values. Each baseline value can be a
        single entity or of the same shape as the model_input (except for the
        batch dimension).
      encoded_baselines: A list of baseline values for the encoded tensor.

    Raises:
      ValueError: If input_name cannot be found in the metadata.
    """
    if input_name not in self._inputs:
      raise ValueError("Input with with name '%s' does not exist." % input_name)
    name = new_name if new_name else input_name
    tensor_name = self._inputs.pop(input_name).input_tensor_name

    self._inputs[name] = explain_metadata.InputMetadata(
        name=name,
        input_tensor_name=tensor_name,
        input_baselines=input_baselines,
        encoded_tensor_name=encoded_name,
        encoded_baselines=encoded_baselines,
        modality=explain_metadata.Modality.CATEGORICAL,
        encoding=encoding)

  def set_image_metadata(
      self,
      input_name: str,
      new_name: Optional[str] = None,
      input_baselines: Optional[List[Union[types.TensorValue,
                                           types.Tensor]]] = None,
      visualization: Optional[Union[Dict[str, str],
                                    parameters.VisualizationParameters]] = None,
      domain: Optional[parameters.DomainInfo] = None) -> None:
    """Sets an existing metadata identified by input as image with params.

    Args:
      input_name: Input name in the metadata to be set as numeric.
      new_name: Optional (unique) new name for this feature.
      input_baselines: A list of baseline values. Each baseline value can be a
        single entity or of the same shape as the model_input (except for the
        batch dimension).
      visualization: Either a dictionary of visualization parameters or
        VisualizationParameters instance. Using VisualizationParameters is
        recommended. If None, a default visualization will be selected based on
        the explanation method (IG/XRAI).
      domain: DomainInfo object specifying the range of the input feature.

    Raises:
      ValueError: If input_name cannot be found in the metadata.
    """
    if input_name not in self._inputs:
      raise ValueError("Input with with name '%s' does not exist." % input_name)
    name = new_name if new_name else input_name
    tensor_name = self._inputs.pop(input_name).input_tensor_name
    if (visualization and
        isinstance(visualization, parameters.VisualizationParameters)):
      visualization = visualization.asdict()
    domain_dict = domain.asdict() if domain else None
    self._inputs[name] = explain_metadata.InputMetadata(
        name=name,
        input_tensor_name=tensor_name,
        input_baselines=input_baselines,
        modality=explain_metadata.Modality.IMAGE,
        visualization=visualization,
        domain=domain_dict)

  def set_output_metadata(self, output_name: str, new_name: str) -> None:
    """Updates an existing output metadata identified by output_name.

    Args:
      output_name: Name of the output that needs to be updated.
      new_name: New (unique) friendly name for the output.

    Raises:
      ValueError: If output with the given name doesn't exist.
    """
    if output_name not in self._outputs:
      raise ValueError("Output with name '%s' does not exist." %
                       output_name)
    if output_name == new_name:
      return
    old_output = self._outputs.pop(output_name)
    self._outputs[new_name] = explain_metadata.OutputMetadata(
        new_name, old_output.output_tensor_name)

  def remove_input_metadata(self, name: str) -> None:
    """Removes input metadata with the name."""
    if name not in self._inputs:
      raise ValueError("Input with with name '%s' does not exist." % name)
    del self._inputs[name]

  def remove_output_metadata(self, name: str) -> None:
    """Removes output metadata with the name."""
    if name not in self._outputs:
      raise ValueError("Output with with name '%s' does not exist." % name)
    del self._outputs[name]

  def get_metadata(self) -> Dict[str, Any]:
    """Returns the current metadata as a dictionary."""
    current_md = explain_metadata.ExplainMetadata(
        inputs=list(self._inputs.values()),
        outputs=list(self._outputs.values()),
        framework=explain_metadata.Framework.TENSORFLOW2,
        tags=[constants.METADATA_TAG])
    return current_md.to_dict()

  def save_metadata(self, file_path: str) -> None:
    """Saves model metadata to the given folder.

    Args:
      file_path: Path to save the model and the metadata. It can be a GCS bucket
        or a local folder. The folder needs to be empty.

    Raises:
      ValueError: If current number of outputs is greater than 1.
    """
    if len(self._outputs) > 1:
      raise ValueError("Number of outputs is more than 1.")
    utils.write_metadata_to_file(self.get_metadata(), file_path)

  def save_model_with_metadata(self, file_path: str) -> str:
    """Saves the model and the generated metadata to the given file path.

    Args:
      file_path: Path to save the model and the metadata. It can be a GCS bucket
        or a local folder. The folder needs to be empty.

    Returns:
      Full file path where the model and the metadata are written.
    """
    kwargs = self._saved_model_args.copy()
    sigs = (
        kwargs.pop("signatures")
        if "signatures" in kwargs else self._loaded_model.signatures)
    tf.saved_model.save(
        self._loaded_model, file_path, signatures=sigs, **kwargs)
    self.save_metadata(file_path)
    return file_path