components/optimization_guide/proto/models.proto

// Copyright 2019 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

syntax = "proto2";
option optimize_for = LITE_RUNTIME;
option java_package = "org.chromium.components.optimization_guide.proto";
option java_outer_classname = "ModelsProto";

package optimization_guide.proto;

import "components/optimization_guide/proto/common_types.proto";

// A generic handle for any type of model.
message Model {
  reserved 3, 4;

  oneof model {
    DecisionTree decision_tree = 1;
    Ensemble ensemble = 2;
    // When passed from the server, this is the URL that the model can be
    // downloaded from. When used internally within Chrome, this contains the
    // absolute file path where the model file is saved on disk.
    string download_url = 5;
  }
  // The tag number is high to allow models to be added and an uncommon number
  // in case the proto this is generated from adds a similar functionality.
  optional DoubleValue threshold = 123;
}

// An ensemble prediction model consisting of an ordered sequence of models.
// This message can be used to express bagged or boosted models.
message Ensemble {
  reserved 2, 3, 4;

  message Member { optional Model submodel = 1; }
  // The tag number is set by the proto this is generated from and cannot be
  // changed.
  repeated Member members = 100;
}

// A decision tree model with its weight for use if included in an ensemble.
message DecisionTree {
  reserved 2;

  repeated TreeNode nodes = 1;
  optional float weight = 3;
}

// A node of a decision tree that is a binary deicison or a leaf.
message TreeNode {
  reserved 6, 7;

  // Following fields are provided for convenience and better readability.
  // Filling them in is not required.
  optional Int32Value node_id = 1;
  optional Int32Value depth = 2;
  optional Int32Value subtree_size = 3;

  oneof node_type {
    BinaryNode binary_node = 4;
    Leaf leaf = 5;
  }
}

// A tree node that contains an inequality test that during evaluation
// determines whether to continue the left or right child.
message BinaryNode {
  reserved 3, 5;

  optional Int32Value left_child_id = 1;
  optional Int32Value right_child_id = 2;
  enum Direction {
    LEFT = 0;
    RIGHT = 1;
  }
  // When a datapoint satisfies the test, it should be propagated to the left
  // child.
  optional InequalityTest inequality_left_child_test = 4;
}

// Vector of values for use within Models.
message Vector {
  repeated Value value = 1;
}

// A leaf node of a decision tree.
message Leaf {
  reserved 2, 3;

  optional Vector vector = 1;
}

// The ID for the features used during evaluation of a Model.
message FeatureId {
  reserved 2;

  optional StringValue id = 1;
}

// The set of inequality operations supported by binary nodes for
// decision tree models.
message InequalityTest {
  reserved 4;

  // When the feature is missing, the test's outcome is undefined.
  optional FeatureId feature_id = 1;
  enum Type {
    LESS_OR_EQUAL = 0;
    LESS_THAN = 1;
    GREATER_OR_EQUAL = 2;
    GREATER_THAN = 3;
  };
  optional Type type = 2;
  optional Value threshold = 3;
}

// Represents a single value of any type, e.g. 5 or "abc".
message Value {
  reserved 5;

  oneof value {
    float float_value = 1;
    double double_value = 2;
    int32 int32_value = 3;
    int64 int64_value = 4;
  }
}

// Wrapper message for `int32`.
//
// The JSON representation for `Int32Value` is JSON number.
message Int32Value {
  // The int32 value.
  optional int32 value = 1;
}

// Wrapper message for `double`.
//
// The JSON representation for `DoubleValue` is JSON number.
message DoubleValue {
  // The double value.
  optional double value = 1;
}

// Requests prediction models to be used for a set of optimization targets.
message GetModelsRequest {
  reserved 2, 4;

  // Information about the requested models.
  repeated ModelInfo requested_models = 1;
  // Context in which this request is made.
  //
  // If the context matches one that requires more urgency (i.e.
  // CONTEXT_PAGE_NAVIGATION), then no model updates will be returned for the
  // requested models.
  optional RequestContext request_context = 3;
  // The locale to associate with this request.
  //
  // It is the IETF language tag, defined in BCP 47. The region subtag is not
  // included when it adds no distinguishing information to the language tag
  // (e.g. both "en-US" and "fr" are correct here).
  optional string locale = 5;
}

// Response to the GetModels request.
message GetModelsResponse {
  // The models to be used during prediction for the requested optimization
  // targets.
  repeated PredictionModel models = 1;
  // A set of model features and their values for the hosts contained in the
  // request to be expected to be consulted with during prediction.
  //
  // It is not guaranteed that this set will contain an entry for every
  // requested host.
  repeated HostModelFeatures host_model_features = 2;
}

// Holds the prediction model for a particular optimization target.
message PredictionModel {
  // Information about the model.
  optional ModelInfo model_info = 1;
  // The model to evaluate for the attached model information.
  //
  // This will only be set if the model that the client claims it has is stale.
  // It is also guaranteed that the value populated as part of this field is one
  // that the client claims to support based on the request's client model
  // capabilities.
  optional Model model = 2;
}

message AdditionalModelFile {
  // When sent by the server, this contains the basenames of the additional
  // files that should be kept and sent to this model's consumers. When used
  // only locally within Chrome, the full path is given.
  optional string file_path = 1;
}

// Metadata for a prediction model for a specific optimization target.
//
// Next ID: 10
message ModelInfo {
  reserved 3;

  // The optimization target for which the model predicts.
  optional OptimizationTarget optimization_target = 1;
  // The version of the model, which is specific to the optimization target.
  optional int64 version = 2;
  // The set of model engine versions the requesting client can use to do model
  // inference.
  repeated ModelEngineVersion supported_model_engine_versions = 4;
  // The set of host model features that are referenced by the model.
  //
  // Note that this should only be populated if part of the response.
  repeated string supported_host_model_features = 5;
  // Additional files required by this model version.
  //
  // If populated, these files are included in the downloaded archive for this
  // model and should be passed along to the consumer.
  //
  // This does not need to be sent to the server in the request for an update to
  // this model. The server will ignore this if sent.
  repeated AdditionalModelFile additional_files = 7;
  // How long the model will remain valid in client storage. If
  // |keep_beyond_valid_duration| is true, will be ignored.
  optional Duration valid_duration = 8;
  // Whether to delete the model once valid_duration has passed.
  optional bool keep_beyond_valid_duration = 9;
  // Mechanism used for model owners to attach metadata to the request or
  // response.
  //
  // In practice, we expect this to be used as a way to negotiate capabilities.
  // The client can provide the model features they can evaluate if this field
  // is part of the request, and the server can provide the model features that
  // are actually present in the model.
  optional Any model_metadata = 6;
}

// The scenarios for which the optimization guide has models for.
enum OptimizationTarget {
  OPTIMIZATION_TARGET_UNKNOWN = 0;
  // Should only be applied when the page load is predicted to be painful.
  OPTIMIZATION_TARGET_PAINFUL_PAGE_LOAD = 1;
  // Target for supplying the language detection model via the model downloader.
  OPTIMIZATION_TARGET_LANGUAGE_DETECTION = 2;
  // Target for determining topics present on a page.
  OPTIMIZATION_TARGET_PAGE_TOPICS = 3;
  // Target for segmentation: New tab page user.
  OPTIMIZATION_TARGET_SEGMENTATION_NEW_TAB = 4;
  // Target for segmentation: Share user.
  OPTIMIZATION_TARGET_SEGMENTATION_SHARE = 5;
  // Target for segmentation: Voice user.
  OPTIMIZATION_TARGET_SEGMENTATION_VOICE = 6;
  // Target for model validation.
  OPTIMIZATION_TARGET_MODEL_VALIDATION = 7;
  // Target for determining entities present on a page.
  OPTIMIZATION_TARGET_PAGE_ENTITIES = 8;
  // Target for Chrome Permissions Suggestions Service: Notification permission.
  OPTIMIZATION_TARGET_NOTIFICATION_PERMISSION_PREDICTIONS = 9;
  // Target that enables data collection on client side for various experiments.
  OPTIMIZATION_TARGET_SEGMENTATION_DUMMY = 10;
  // Target for segmentation: Chrome Android Start user.
  OPTIMIZATION_TARGET_SEGMENTATION_CHROME_START_ANDROID = 11;
  // Target for segmentation: Query Tiles user.
  OPTIMIZATION_TARGET_SEGMENTATION_QUERY_TILES = 12;
  // Target for determining the UI visibility of a page.
  OPTIMIZATION_TARGET_PAGE_VISIBILITY = 13;
  // Target for supplying the Autofill Assistant annotate DOM model via the
  // model downloader.
  OPTIMIZATION_TARGET_AUTOFILL_ASSISTANT = 14;
  // Target for determining topics present on a page.
  // TODO(crbug/1266504): Remove PAGE_TOPICS in favor of this target.
  OPTIMIZATION_TARGET_PAGE_TOPICS_V2 = 15;
  // Target for segmentation: Determine users with low engagement with chrome.
  OPTIMIZATION_TARGET_SEGMENTATION_CHROME_LOW_USER_ENGAGEMENT = 16;
  // Target for segmentation: Determine users who prefer to use Feed.
  OPTIMIZATION_TARGET_SEGMENTATION_FEED_USER = 17;
  // Target for segmentation: Determine whether price tracking should be shown
  // as a contextual page action.
  OPTIMIZATION_TARGET_CONTEXTUAL_PAGE_ACTION_PRICE_TRACKING = 18;
  // Target for smart text selection and entity extraction.
  OPTIMIZATION_TARGET_TEXT_CLASSIFIER = 19;
}

// The model engine versions that can be used to do model inference.
//
// Please only update these enums when a new major version of TFLite rolls.
//
// For example: v1.2.3
//                 ^
//                 Change when this number increments.
enum ModelEngineVersion {
  MODEL_ENGINE_VERSION_UNKNOWN = 0;
  // A decision tree.
  MODEL_ENGINE_VERSION_DECISION_TREE = 1;
  // A model using only operations that are supported by TensorflowLite 2.3.0.
  MODEL_ENGINE_VERSION_TFLITE_2_3_0 = 2;
  // A model using only operations that are supported by TensorflowLite 2.3.0
  // with updated FULLY_CONNECTED and BATCH_MUL versions for quantized models.
  MODEL_ENGINE_VERSION_TFLITE_2_3_0_1 = 3;
  // TensorflowLite version 2.4.2, and a bit more up to internal rev number
  // 381280669.
  MODEL_ENGINE_VERSION_TFLITE_2_4 = 4;
  // TensorflowLite version 2.7.*. This is where regular ~HEAD rolls started.
  MODEL_ENGINE_VERSION_TFLITE_2_7 = 5;
  // A model using only operations that are supported by TensorflowLite 2.8.0.
  MODEL_ENGINE_VERSION_TFLITE_2_8 = 6;
  // A model using only operations that are supported by TensorflowLite 2.9.0.
  MODEL_ENGINE_VERSION_TFLITE_2_9 = 7;
  // A model using only operations that are supported by TensorflowLite 2.9.0.
  // This adds GELU to the supported ops in Optimiziation Guide.
  MODEL_ENGINE_VERSION_TFLITE_2_9_0_1 = 8;
  // A model using only operations that are supported by TensorflowLite 2.10.0.
  MODEL_ENGINE_VERSION_TFLITE_2_10 = 9;
}

// A set of model features and the host that it applies to.
message HostModelFeatures {
  // The host that the features should be applied for.
  optional string host = 1;
  // The set of features and their values that apply to the host.
  repeated ModelFeature model_features = 2;
}

// Information about a feature that is potentially referenced in a model.
message ModelFeature {
  // The name of the feature to match if encountered in a model.
  optional string feature_name = 1;
  // The value of the feature to be used during prediction.
  oneof feature_value {
    double double_value = 2;
    int64 int64_value = 3;
  }
}