Configure example-based explanations

To use example-based explanations, you must configure explanations by specifying an explanationSpec when you import or upload the Model resource into the Model Registry.

Then, when you request online explanations, you can override some of those configuration values by specifying an ExplanationSpecOverride in the request. You cannot request batch explanations; they are not supported.

This page describes how to configure and update these options.

Configure explanations when importing or uploading the model

Before you begin, make sure that you have following:

  1. A Cloud Storage location that contains your model artifacts. Your model either needs to be a deep neural network (DNN) model where you provide the name of a layer, or signature, whose output can be used as the latent space, or you can provide a model that directly outputs embeddings (latent space representation). This latent space captures the example representations that are used for generating explanations.

  2. A Cloud Storage location that contains the instances to be indexed for approximate nearest neighbor search. For more information, see input data requirements.

Console

Follow the guide to importing a model using the Google Cloud console.

In the Explainability tab, select Example-based explanation and fill in the fields.

For information about each field, see the tips in Google Cloud Console (shown below) as well as the reference documentation for Example and ExplanationMetadata.

gcloud CLI

  1. Write the following ExplanationMetadata to a JSON file in your local environment. The filename does not matter, but for this example, call the file explanation-metadata.json:
{
  "inputs": {
    "my_input": {
      "inputTensorName": "INPUT_TENSOR_NAME",
      "encoding": "IDENTITY",
    },
    "id": {
      "inputTensorName": "id",
      "encoding": "IDENTITY"
    }
  },
  "outputs": {
    "embedding": {
      "outputTensorName": "OUTPUT_TENSOR_NAME"
    }
  }
}
  1. (Optional) If you are specifying the full NearestNeighborSearchConfig, write the following to a JSON file in your local environment. The filename does not matter, but for this example, call the file search_config.json:
{
  "contentsDeltaUri": "",
  "config": {
      "dimensions": 50,
      "approximateNeighborsCount": 10,
      "distanceMeasureType": "SQUARED_L2_DISTANCE",
      "featureNormType": "NONE",
      "algorithmConfig": {
          "treeAhConfig": {
              "leafNodeEmbeddingCount": 1000,
              "fractionLeafNodesToSearch": 1.0
          }
      }
    }
  }
  1. Run the following command to upload your Model.

If you are using a Preset search configuration, remove the --explanation-nearest-neighbor-search-config-file flag. If you are specifying NearestNeighborSearchConfig, remove the --explanation-modality and --explanation-query flags.

The flags most pertinent to example-based explanations are bolded.

gcloud ai models upload \
    --region=LOCATION \
    --display-name=MODEL_NAME \
    --container-image-uri=IMAGE_URI \
    --artifact-uri=MODEL_ARTIFACT_URI \
    --explanation-method=examples \
    --uris=[URI, ...] \
    --explanation-neighbor-count=NEIGHBOR_COUNT \
    --explanation-metadata-file=explanation-metadata.json \
    --explanation-modality=IMAGE|TEXT|TABULAR \
    --explanation-query=PRECISE|FAST \
    --explanation-nearest-neighbor-search-config-file=search_config.json

See gcloud ai models upload for more information.

  1. The upload action returns an OPERATION_ID that can be used to check when the operation is finished. You can poll for the status of the operation until the response includes "done": true. Use the gcloud ai operations describe command to poll the status, for example:

    gcloud ai operations describe <operation-id>
    

    You will not be able to request explanations until the operation is done. Depending on the size of the dataset and model architecture, this step can take several hours to build the index used to query for examples.

REST

Before using any of the request data, make the following replacements:

  • PROJECT
  • LOCATION

To learn about the other placeholders, see Model, explanationSpec, and Examples.

To learn more about uploading models, see the upload method and Importing models.

The JSON request body below specifies a Preset search configuration. Alternatively, you can specify the fullNearestNeighborSearchConfig.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models:upload

Request JSON body:

{
  "model": {
    "displayName": "my-model",
    "artifactUri": "gs://your-model-artifact-folder",
    "containerSpec": {
      "imageUri": "us-docker.pkg.dev/vertex-ai/prediction/tf2-cpu.2-11:latest",
    },
    "explanationSpec": {
      "parameters": {
        "examples": {
          "gcsSource": {
            "uris": ["gs://your-examples-folder"]
          },
          "neighborCount": 10,
          "presets": {
            "modality": "image"
          }
        }
      },
      "metadata": {
        "outputs": {
            "embedding": {
                "output_tensor_name": "embedding"
            }
        },
        "inputs": {
          "my_fancy_input": {
            "input_tensor_name": "input_tensor_name",
            "encoding": "identity",
            "modality": "image"
          },
          "id": {
            "input_tensor_name": "id",
            "encoding": "identity"
          }
        }
      }
    }
  }
}

To send your request, expand one of these options:

You should receive a JSON response similar to the following:

{
  "name": "projects/PROJECT_NUMBER/locations/LOCATION/models/MODEL_ID/operations/OPERATION_ID",
  "metadata": {
    "@type": "type.googleapis.com/google.cloud.aiplatform.v1.UploadModelOperationMetadata",
    "genericMetadata": {
      "createTime": "2022-01-08T01:21:10.147035Z",
      "updateTime": "2022-01-08T01:21:10.147035Z"
    }
  }
}

The upload action returns an OPERATION_ID that can be used to check when the operation is finished. You can poll for the status of the operation until the response includes "done": true. Use the gcloud ai operations describe command to poll the status, for example:

gcloud ai operations describe <operation-id>

You will not be able to request explanations until the operation is done. Depending on the size of the dataset and model architecture, this step can take several hours to build the index used to query for examples.

Python

See the section Upload the model in the image classification example-based explanations notebook.

NearestNeighborSearchConfig

The following JSON request body demonstrates how to specify the full NearestNeighborSearchConfig (instead of presets) in an upload request.

{
  "model": {
    "displayName": displayname,
    "artifactUri": model_path_to_deploy,
    "containerSpec": {
      "imageUri": DEPLOY_IMAGE,
    },
    "explanationSpec": {
      "parameters": {
        "examples": {
          "gcsSource": {
            "uris": [DATASET_PATH]
          },
          "neighborCount": 5,
          "nearest_neighbor_search_config": {
            "contentsDeltaUri": "",
            "config": {
              "dimensions": dimensions,
              "approximateNeighborsCount": 10,
              "distanceMeasureType": "SQUARED_L2_DISTANCE",
              "featureNormType": "NONE",
              "algorithmConfig": {
                  "treeAhConfig": {
                      "leafNodeEmbeddingCount": 1000,
                      "fractionLeafNodesToSearch": 1.0
                  }
                }
              }
          }
        }
      },
      "metadata": { ... }
    }
  }
}

The tables below lists the fields for NearestNeighborSearchConfig.

Fields
dimensions

int32

Required. The number of dimensions of the input vectors. Used for dense embeddings only.

approximateNeighborsCount

int32

Required if tree-AH algorithm is used.

The default number of neighbors to find through approximate search before exact reordering is performed. Exact reordering is a procedure where results returned by an approximate search algorithm are reordered using a more expensive distance computation.

ShardSize ShardSize

The size of each shard. When an index is large, it is sharded based on the specified shard size. During serving, each shard is served on a separate node and scales independently.

distanceMeasureType

DistanceMeasureType

The distance measure used in nearest neighbor search.

featureNormType

FeatureNormType

Type of normalization to be carried out on each vector.

algorithmConfig oneOf:

The configuration for the algorithms that Vector Search uses for efficient search. Used for dense embeddings only.

  • TreeAhConfig: Configuration options for using the tree-AH algorithm. For more information, see this blog Scaling deep retrieval with TensorFlow Recommenders and Vector Search
  • BruteForceConfig: This option implements the standard linear search in the database for each query. There are no fields to configure for a brute force search. To select this algorithm, pass an empty object for BruteForceConfig.

DistanceMeasureType

Enums
SQUARED_L2_DISTANCE Euclidean (L2) Distance
L1_DISTANCE Manhattan (L1) Distance
DOT_PRODUCT_DISTANCE Default value. Defined as a negative of the dot product.
COSINE_DISTANCE Cosine Distance. We strongly suggest using DOT_PRODUCT_DISTANCE + UNIT_L2_NORM instead of the COSINE distance. Our algorithms have been more optimized for the DOT_PRODUCT distance, and when combined with UNIT_L2_NORM, it offers the same ranking and mathematical equivalence as the COSINE distance.

FeatureNormType

Enums
UNIT_L2_NORM Unit L2 normalization type.
NONE Default value. No normalization type is specified.

TreeAhConfig

These are the fields to select for the tree-AH algorithm (Shallow tree + Asymmetric Hashing).

Fields
fractionLeafNodesToSearch double
The default fraction of leaf nodes that any query may be searched. Must be in range 0.0 - 1.0, exclusive. The default value is 0.05 if not set.
leafNodeEmbeddingCount int32
Number of embeddings on each leaf node. The default value is 1000 if not set.
leafNodesToSearchPercent int32
Deprecated, use fractionLeafNodesToSearch.

The default percentage of leaf nodes that any query may be searched. Must be in range 1-100, inclusive. The default value is 10 (means 10%) if not set.

BruteForceConfig

This option implements the standard linear search in the database for each query. There are no fields to configure for a brute force search. To select this algorithm, pass an empty object for BruteForceConfig to algorithmConfig.

Input data requirements

Upload your dataset to a Cloud Storage location. Make sure the files in the JSON Lines format.

The files must be in JSON Lines format. The following sample is from the image classification example-based explanations notebook:

{"id": "0", "bytes_inputs": {"b64": "..."}}
{"id": "1", "bytes_inputs": {"b64": "..."}}
{"id": "2", "bytes_inputs": {"b64": "..."}}

Update the index or configuration

Vertex AI lets you update a model's nearest neighbor index or Example configuration. This is useful if you'd like to update your model without re-indexing its dataset. For example, if your model's index contains 1,000 instances, and you'd like to add 500 more instances, you can call UpdateExplanationDataset to add to the index without re-processing the original 1,000 instances.

To update the explanation dataset:

Python

def update_explanation_dataset(model_id, new_examples):
    response = clients["model"].update_explanation_dataset(model=model_id,  examples=new_examples)
    update_dataset_response = response.result()
    return update_dataset_response

PRESET_CONFIG = {
    "modality": "TEXT",
    "query": "FAST"
}
NEW_DATASET_FILE_PATH = "new_dataset_path"
NUM_NEIGHBORS_TO_RETURN = 10

EXAMPLES = aiplatform.Examples(presets=PRESET_CONFIG,
                                    gcs_source=aiplatform.types.io.GcsSource(uris=[NEW_DATASET_FILE_PATH]),
                                      neighbor_count=NUM_NEIGHBORS_TO_RETURN)

MODEL_ID = 'model_id'
update_dataset_response = update_explanation_dataset(MODEL_ID, EXAMPLES)

Usage notes:

  • The model_id remains unchanged after the UpdateExplanationDataset operation.

  • The UpdateExplanationDataset operation affects only the Model resource; it does not update any associated DeployedModels. This means that a deployedModel's index contains the dataset at the time it was deployed. To update a deployedModel's index, you must re-deploy the updated model to an endpoint.

Override the configuration when getting online explanations

When you request an explanation, you can override some of the parameters on the fly by specifying the ExplanationSpecOverride field.

Depending on the application, some constraints might be desirable on the kind of explanations that are returned. For example, to ensure diversity of explanations, a user can specify a crowding parameter which dictates that no single type of examples are over-represented in the explanations. Concretely, if a user is trying to understand why a bird was labeled as a plane by their model, they might not be interested in seeing too many bird examples as explanations to better investigate the root cause.

The following table summarizes the parameters that can be overridden for an example-based explanation request:

Property Name Property Value Description
neighborCount int32 The number of examples to return as explanation
crowdingCount int32 Maximum number of examples to return with the same crowding tag
allow String Array The tags that are allowed for explanations to have
deny String Array The tags that are not allowed for explanations to have

The Vector Search Filtering describes these parameters in more details.

Here's an example of a JSON request body with overrides:

{
  "instances":[
    {
      "id": data[0]["id"],
      "bytes_inputs": {"b64": bytes},
      "restricts": "",
      "crowding_tag": ""
    }
  ],
  "explanation_spec_override": {
    "examples_override": {
      "neighbor_count": 5,
      "crowding_count": 2,
      "restrictions": [
        {
          "namespace_name": "label",
          "allow": ["Papilloma", "Rift_Valley", "TBE", "Influenza", "Ebol"]
        }
      ]
    }
  }
}

What's next

Here's an example of the response from an example-based explain request:

[
   {
      "neighbors":[
         {
            "neighborId":"311",
            "neighborDistance":383.8
         },
         {
            "neighborId":"286",
            "neighborDistance":431.4
         }
      ],
      "input":"0"
   },
   {
      "neighbors":[
         {
            "neighborId":"223",
            "neighborDistance":471.6
         },
         {
            "neighborId":"55",
            "neighborDistance":392.7
         }
      ],
      "input":"1"
   }
]

Pricing

See the section on example-based explanations in the pricing page.