Method: projects.explain

Performs explanation on the data in the request.

AI Explanations implements a custom explain verb on top of an HTTP POST method. The explain method performs prediction on the data in the request.

The URL is described in Google API HTTP annotation syntax:

POST https://ml.googleapis.com/v1/{name=projects/**}:explain

The name parameter is required. It must contain the name of your model and, optionally, a version. If you specify a model without a version, the default version for that model is used.

Example specifying both model and version:

POST https://ml.googleapis.com/v1/projects/YOUR_PROJECT_ID/models/YOUR_MODEL_NAME/versions/YOUR_VERSION_NAME:explain

Example specifying only a model. The default version for that model is used:

POST https://ml.googleapis.com/v1/projects/YOUR_PROJECT_ID/models/YOUR_MODEL_NAME:explain

This page describes the format of the explanation request body and of the response body. For a code sample showing how to send an explanation request, see the guide to using feature attributions.

Request body details

TensorFlow

The request body contains data with the following structure (JSON representation):

{
  "instances": [
    <value>|<simple/nested list>|<object>,
    ...
  ]
}

The instances[] object is required, and must contain the list of instances to get explanations for.

The structure of each element of the instances list is determined by your model's input definition. Instances can include named inputs (as objects) or can contain only unlabeled values.

Not all data includes named inputs. Some instances are simple JSON values (boolean, number, or string). However, instances are often lists of simple values, or complex nested lists.

Below are some examples of request bodies.

CSV data with each row encoded as a string value:

{"instances": ["1.0,true,\\"x\\"", "-2.0,false,\\"y\\""]}

Plain text:

{"instances": ["the quick brown fox", "the lazy dog"]}

Sentences encoded as lists of words (vectors of strings):

{
  "instances": [
    ["the","quick","brown"],
    ["the","lazy","dog"],
    ...
  ]
}

Floating point scalar values:

{"instances": [0.0, 1.1, 2.2]}

Vectors of integers:

{
  "instances": [
    [0, 1, 2],
    [3, 4, 5],
    ...
  ]
}

Tensors (in this case, two-dimensional tensors):

{
  "instances": [
    [
      [0, 1, 2],
      [3, 4, 5]
    ],
    ...
  ]
}

Images, which can be represented different ways. In this encoding scheme the first two dimensions represent the rows and columns of the image, and the third dimension contains lists (vectors) of the R, G, and B values for each pixel:

{
  "instances": [
    [
      [
        [138, 30, 66],
        [130, 20, 56],
        ...
      ],
      [
        [126, 38, 61],
        [122, 24, 57],
        ...
      ],
      ...
    ],
    ...
  ]
}

Data encoding

JSON strings must be encoded as UTF-8. To send binary data, you must base64-encode the data and mark it as binary. To mark a JSON string as binary, replace it with a JSON object with a single attribute named b64:

{"b64": "..."} 

The following example shows two serialized tf.Examples instances, requiring base64 encoding (fake data, for illustrative purposes only):

{"instances": [{"b64": "X5ad6u"}, {"b64": "IA9j4nx"}]}

The following example shows two JPEG image byte strings, requiring base64 encoding (fake data, for illustrative purposes only):

{"instances": [{"b64": "ASa8asdf"}, {"b64": "JLK7ljk3"}]}

Multiple input tensors

Some models have an underlying TensorFlow graph that accepts multiple input tensors. In this case, use the names of JSON name/value pairs to identify the input tensors.

For a graph with input tensor aliases "tag" (string) and "image" (base64-encoded string):

{
  "instances": [
    {
      "tag": "beach",
      "image": {"b64": "ASa8asdf"}
    },
    {
      "tag": "car",
      "image": {"b64": "JLK7ljk3"}
    }
  ]
}

For a graph with input tensor aliases "tag" (string) and "image" (3-dimensional array of 8-bit ints):

{
  "instances": [
    {
      "tag": "beach",
      "image": [
        [
          [138, 30, 66],
          [130, 20, 56],
          ...
        ],
        [
          [126, 38, 61],
          [122, 24, 57],
          ...
        ],
        ...
      ]
    },
    {
      "tag": "car",
      "image": [
        [
          [255, 0, 102],
          [255, 0, 97],
          ...
        ],
        [
          [254, 1, 101],
          [254, 2, 93],
          ...
        ],
        ...
      ]
    },
    ...
  ]
}

Explanation metadata

When using AI Explanations, you need to indicate which of your tensors correspond to your actual features and output probabilities or predictions. You do this by adding a file named explanation_metadata.json to your SavedModel folder before deploying the model to AI Explanations.

To make this process easier, assign a name to the tensors in your TensorFlow graph. Before training your model, set the name property in either raw tensors or Keras layers:

auxiliary_output = Dense(1, activation='sigmoid', name='aux_output')

The file contents should match this schema:

{
  "inputs": {
    string <input feature key>: {
      "input_tensor_name": string,
      "input_baselines": [
        number,
      ],
      "modality": string
    }
    ...
  },
  "outputs": {
    string <output value key>:  {
      "output_tensor_name": string
    },
    ...
  },
  "framework": string
}
Fields

output value key and input feature key

Any unique name. The system outputs a dictionary with the attribution scores for a given feature listed under this key.
input_tensor_name

string

Required. The name of the tensor containing the inputs that the model's prediction should be attributed to. Format the name as name:0. For example, aux_output:0.

input_baselines

<integer>|<simple list>

Optional. The value of the baselines or "uninformative" example for this particular feature. Consider using the average or 0.
The shape of each baseline should match the shape of the encoded tensor. If a scalar is provided, we broadcast to the same shape as the encoded tensor.
If you supply multiple baselines, the system averages the attributions among the baselines. For example, you might want to compare attributions to either a fully black or white image.

modality

string

Optional. Can be set to image if the input tensor is an image. In that case, the system will return a graphical representation of the attributions.

The tensor specified by input_tensor_name should be:

  • For color images: A dense 4-D tensor of dtype float32 and shape [batch_size, height, width, 3] whose elements are RGB color values of pixels normalized to the range [0, 1].
  • For grayscale images: A dense 3-D tensor of dtype float32 and shape [batch_size, height, width] whose elements are black values of pixels normalized to the range [0, 1].
output_tensor_name

string

Required. The name of the tensor containing the outputs that the model's prediction should be attributed to.

framework

string

Required. Must be set to tensorflow.

Configure visualization settings

When you get explanations on image data with the integrated gradients or XRAI methods, you can configure visualization settings for your results by including them in your explanation_metadata.json file. Configuring these settings is optional.

To configure your visualiztion settings, include the visualization config within the input object you want to visualize:

{
  "inputs": {
    string <input feature key>: {
      "input_tensor_name": string,
      "input_baselines": [
        number,
      ],
      "modality": string
      "visualization": {
        "type": string,
        "polarity": string,
        "clip_below_percentile": number,
        "clip_above_percentile": number,
        "color_map": string,
        "overlay_type": string
      }
    }
    ...
  },
  "outputs": {
    string <output value key>:  {
      "output_tensor_name": string
    },
    ...
  },
  "framework": string
}

Details on visualization settings

All of the visualization settings are optional, so you can specify all, some, or none of these values.

"visualization": {
  "type": string,
  "polarity": string,
  "clip_below_percentile": number,
  "clip_above_percentile": number,
  "color_map": string,
  "overlay_type": string
}

See example configurations and output images.

Fields

type

string

Optional. The type of visualization. Valid values are outlines or pixels. For integrated gradients, you can use either setting. The default setting is outlines.
For XRAI, pixels is the default setting. outlines is not recommended for XRAI.

polarity

string

Optional. The directionality of the attribution values displayed. Valid values are positive, negative, or both. Defaults to positive.

clip_below_percentile

number

Optional.. Excludes attributions below the specified percentile. Valid value is a decimal in the range [0, 100].

clip_above_percentile

number

Optional. Excludes attributions above the specified percentile. Valid value is a decimal in the range [0, 100]. Must be larger than clip_below_percentile.

color_map

string

Optional. Valid values are red_green, pink_green, and viridis. Defaults to pink_green for integrated gradients and viridis for XRAI.

overlay_type

string

Optional. The type of overlay modifying how the attributions are displayed over the original input images.
Valid values are none, grayscale, original and mask_black.

  • none: The attributions are displayed alone over a black image, without being overlaid onto the input image.
  • grayscale: The attributions are overlaid onto a grayscaled version of the input image.
  • original: The attributions are overlaid onto the original input image.
  • mask_black: The attributions are used as a mask to emphasize predictive parts of the image. The opacity of the pixels in the original image correspond to the intensity of the attributions for the corresponding pixel.

Defaults to original for integrated gradients and grayscale for XRAI.

See example images showing each overlay_type.

Response body details

Responses are very similar to requests.

If the call is successful, the response body contains one explanations entry per instance in the request body, given in the same order:

{
  "explanations": [
    {
      object
    }
  ]
}

If explanation fails for any instance, the response body contains no explanations. Instead, it contains a single error entry:

{
  "error": string
}

The explanations[] object contains the list of explanations, one for each instance in the request.

On error, the error string contains a message describing the problem. The error is returned instead of an explanations list if an error occurred while processing any instance.

Even though there is one explanation per instance, the format of an explanation is not directly related to the format of an instance. Explanations take whatever format is specified in the outputs collection defined in the model. The collection of explanations is returned in a JSON list. Each member of the list can be a simple value, a list, or a JSON object of any complexity. If your model has more than one output tensor, each explanation will be a JSON object containing a name/value pair for each output. The names identify the output aliases in the graph.

Response body examples

The following explanations response is for an individual feature attribution on tabular data. It is part of the example notebook for tabular data. The notebook demonstrates how to parse explanations responses and plot the attributions data.

Feature attributions appear within the attributions_by_label object:

{
 "explanations": [
  {
   "attributions_by_label": [
    {
     "approx_error": 0.001017811509478243,
     "attributions": {
      "data": [
       -0.0,
       1.501250445842743,
       4.4058547498107075,
       0.016078486742916454,
       -0.03749384209513669,
       -0.0,
       -0.2621846305120581,
       -0.0,
       -0.0,
       0.0
      ]
      ...
     },
     "baseline_score": 14.049912452697754,
     "example_score": 19.667699813842773,
     "label_index": 0,
     "output_name": "duration"
    }
    ...
   ]
  }
 ]
}
  • approx_error is an approximation error for the feature attributions. Feature attributions are based on an approximation of Shapley values. Learn more about the approximation error.
  • The attributions object contains a key-value pair for each input feature you requested explanations for.
    • For each input feature, the key is the same as the input feature key you set in your explanation_metadata.json file. In this example, it's "data".
    • The values are the attributions for each feature. The shape of the attributions matches the input tensor. In this example, it is a list of scalar values. Learn more about these attribution values.
  • The baseline_score is the model output for the baseline you set. Depending on your model, you can set the baseline to zero, a random value, or median values. Learn more about selecting baselines.
  • The example_score is the prediction for the input instance you provided. In this example, example_score is the predicted duration of a rideshare bike trip in minutes. In general, the example_score for any given instance is the prediction for that instance.
  • The label_index is the index in the output tensor that is being explained.
  • The output_name is the same as the output feature key you set in your explanation_metadata.json file. In this example, it's "duration".

Individual attribution values

This table shows more details about the features that correspond to these example attribution values. Positive attribution values increase the predicted value by that amount, and negative attribution values decrease the predicted value. The euclidean distance between the start and end locations of the bike trip had the strongest effect on the predicted bike trip duration (19.667 minutes), increasing it by 4.498.

Feature name Feature value Attribution value
start_hr 19 0
weekday 1 -0.0661425
euclidean 3920.76 4.49809
temp 52.5 0.0564195
dew_point 38.8 -0.072438
wdsp 0 0
max_temp 64.8 -0.226125
fog 0 -0
prcp 0.06 -0
rain_drizzle 0 0

Learn more by trying the example notebook for tabular data.

API specification

The following section describes the specification of the explain method as defined in the AI Platform Training and Prediction API discovery document. Refer to the previous sections of this document for detailed information about the method.

HTTP request

POST https://{endpoint}/v1/{name=projects/**}:explain

Where {endpoint} is one of the supported service endpoints.

The URLs use gRPC Transcoding syntax.

Path parameters

Parameters
name

string

Required. The resource name of a model or a version.

Authorization: requires the predict permission on the specified resource.

Authorization requires one or more of the following IAM permissions on the specified resource name:

  • ml.models.predict
  • ml.versions.predict

Request body

The request body contains data with the following structure:

JSON representation
{
  "httpBody": {
    object (HttpBody)
  }
}
Fields
httpBody

object (HttpBody)

Required. The explanation request body.

Response body

If successful, the response is a generic HTTP response whose format is defined by the method.

Authorization Scopes

Requires the following OAuth scope:

  • https://www.googleapis.com/auth/cloud-platform

For more information, see the Authentication Overview.