android-12.0.0_r34/s

/* Copyright 2019 The TensorFlow Authors. 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.
==============================================================================*/

syntax = "proto2";

package tflite.evaluation;

import "tensorflow/lite/tools/evaluation/proto/preprocessing_steps.proto";

option cc_enable_arenas = true;
option java_multiple_files = true;
option java_package = "tflite.evaluation";

// Defines the functionality executed by an EvaluationStage.
//
// Next ID: 7
message ProcessSpecification {
  oneof params {
    ImagePreprocessingParams image_preprocessing_params = 1;
    TopkAccuracyEvalParams topk_accuracy_eval_params = 2;
    TfliteInferenceParams tflite_inference_params = 3;
    ImageClassificationParams image_classification_params = 4;
    ObjectDetectionAveragePrecisionParams
        object_detection_average_precision_params = 5;
    ObjectDetectionParams object_detection_params = 6;
  }
}

// Latency numbers in microseconds, based on all EvaluationStage::Run() calls so
// far.
//
// Next ID: 7
message LatencyMetrics {
  // Latency for the last Run.
  optional int64 last_us = 1;
  // Maximum latency observed for any Run.
  optional int64 max_us = 2;
  // Minimum latency observed for any Run.
  optional int64 min_us = 3;
  // Sum of all Run latencies.
  optional int64 sum_us = 4;
  // Average latency across all Runs.
  optional double avg_us = 5;
  // Standard deviation for latency across all Runs.
  optional int64 std_deviation_us = 6;
}

// Statistics for an accuracy value over multiple runs of evaluation.
//
// Next ID: 5
message AccuracyMetrics {
  // Maximum value observed for any Run.
  optional float max_value = 1;
  // Minimum value observed for any Run.
  optional float min_value = 2;
  // Average value across all Runs.
  optional double avg_value = 3;
  // Standard deviation across all Runs.
  optional float std_deviation = 4;
}

// Contains process-specific metrics, which may differ based on what an
// EvaluationStage does.
//
// Next ID: 8
message ProcessMetrics {
  optional LatencyMetrics total_latency = 1;

  oneof stage_metrics {
    TopkAccuracyEvalMetrics topk_accuracy_metrics = 2;
    TfliteInferenceMetrics tflite_inference_metrics = 3;
    ImageClassificationMetrics image_classification_metrics = 4;
    InferenceProfilerMetrics inference_profiler_metrics = 5;
    ObjectDetectionAveragePrecisionMetrics
        object_detection_average_precision_metrics = 6;
    ObjectDetectionMetrics object_detection_metrics = 7;
  }
}

// Parameters that define how images are preprocessed.
//
// Next ID: 3
message ImagePreprocessingParams {
  // Required.
  repeated ImagePreprocessingStepParams steps = 1;
  // Same as tflite::TfLiteType.
  required int32 output_type = 2;
}

// Parameters that control TFLite inference.
//
// Next ID: 5
message TfliteInferenceParams {
  // Required
  optional string model_file_path = 1;

  enum Delegate {
    NONE = 0;
    NNAPI = 1;
    GPU = 2;
    HEXAGON = 3;
    XNNPACK = 4;
  }
  optional Delegate delegate = 2;
  // Number of threads available to the TFLite Interpreter.
  optional int32 num_threads = 3 [default = 1];

  // Defines how many times the TFLite Interpreter is invoked for every input.
  // This helps benchmark cases where extensive pre-processing might not be
  // required for every input.
  optional int32 invocations_per_run = 4 [default = 1];
}

// Metrics specific to TFLite inference.
//
// Next ID: 2
message TfliteInferenceMetrics {
  // Number of times the interpreter is invoked.
  optional int32 num_inferences = 1;
}

// Parameters that define how top-K accuracy is evaluated.
//
// Next ID: 2
message TopkAccuracyEvalParams {
  // Required.
  optional int32 k = 1;
}

// Metrics from top-K accuracy evaluation.
//
// Next ID: 2
message TopkAccuracyEvalMetrics {
  // A repeated field of size |k| where the ith element denotes the fraction of
  // samples for which the correct label was present in the top (i + 1) model
  // outputs.
  // For example, topk_accuracies(1) will contain the fraction of
  // samples for which the model returned the correct label as the top first or
  // second output.
  repeated float topk_accuracies = 1;
}

// Parameters that define how the Image Classification task is evaluated
// end-to-end.
//
// Next ID: 3
message ImageClassificationParams {
  // Required.
  // TfLite model should have 1 input & 1 output tensor.
  // Input shape: {1, image_height, image_width, 3}
  // Output shape: {1, num_total_labels}
  optional TfliteInferenceParams inference_params = 1;

  // Optional.
  // If not set, accuracy evaluation is not performed.
  optional TopkAccuracyEvalParams topk_accuracy_eval_params = 2;
}

// Metrics from evaluation of the image classification task.
//
// Next ID: 5
message ImageClassificationMetrics {
  optional LatencyMetrics pre_processing_latency = 1;
  optional LatencyMetrics inference_latency = 2;
  optional TfliteInferenceMetrics inference_metrics = 3;
  // Not set if topk_accuracy_eval_params was not populated in
  // ImageClassificationParams.
  optional TopkAccuracyEvalMetrics topk_accuracy_metrics = 4;
}

// Metrics computed from comparing TFLite execution in two settings:
// 1. User-defined TfliteInferenceParams (The 'test' setting)
// 2. Default TfliteInferenceParams (The 'reference' setting)
//
// Next ID: 4
message InferenceProfilerMetrics {
  // Latency metrics from Single-thread CPU inference.
  optional LatencyMetrics reference_latency = 1;
  // Latency from TfliteInferenceParams under test.
  optional LatencyMetrics test_latency = 2;
  // For reference & test output vectors {R, T}, the error is computed as:
  // Mean([Abs(R[i] - T[i]) for i in num_elements])
  // output_errors[v] : statistics for the error value of the vth output vector
  //   across all Runs.
  repeated AccuracyMetrics output_errors = 3;
}

// Proto containing information about all the objects (predicted or
// ground-truth) contained in an image.
//
// Next ID: 4
message ObjectDetectionResult {
  // One instance of an object detected in an image.
  // Next ID: 4
  message ObjectInstance {
    // Defines the bounding box for a detected object.
    // Next ID: 5
    message NormalizedBoundingBox {
      // All boundaries defined below are required.
      // Each boundary value should be normalized with respect to the image
      // dimensions. This helps evaluate detections independent of image size.
      // For example, normalized_top = top_boundary / image_height.
      optional float normalized_top = 1;
      optional float normalized_bottom = 2;
      optional float normalized_left = 3;
      optional float normalized_right = 4;
    }

    // Required.
    optional int32 class_id = 1;
    // Required
    optional NormalizedBoundingBox bounding_box = 2;
    // Value in (0, 1.0] denoting confidence in this prediction.
    // Default value of 1.0 for ground-truth data.
    optional float score = 3 [default = 1.0];
  }

  repeated ObjectInstance objects = 1;
  // Filename of the image.
  optional string image_name = 2;
  // Unique id for the image.
  optional int64 image_id = 3;
}

// Proto containing ground-truth ObjectsSets for all images in a COCO validation
// set.
//
// Next ID: 2
message ObjectDetectionGroundTruth {
  repeated ObjectDetectionResult detection_results = 1;
}

// Parameters that define how Average Precision is computed for Object Detection
// task.
// Refer for details: http://cocodataset.org/#detection-eval
//
// Next ID: 4
message ObjectDetectionAveragePrecisionParams {
  // Total object classes. The AP value returned for each IoU threshold is an
  // average over all classes encountered in predicted/ground truth sets.
  optional int32 num_classes = 1;
  // A predicted box matches a ground truth box if and only if
  // IoU between these two are larger than an IoU threshold.
  // AP is computed for all relevant {IoU threshold, class} combinations and
  // averaged to get mAP.
  // If left empty, evaluation is done for all IoU threshods in the range
  // 0.5:0.05:0.95 (min:increment:max).
  repeated float iou_thresholds = 2;
  // AP is computed as the average of maximum precision at (1
  // + num_recall_points) recall levels. E.g., if num_recall_points is 10,
  // recall levels are 0., 0.1, 0.2, ..., 0.9, 1.0.
  // Default: 100
  optional int32 num_recall_points = 3 [default = 100];
}

// Average Precision metrics from Object Detection task.
//
// Next ID: 3
message ObjectDetectionAveragePrecisionMetrics {
  // Average Precision value for a particular IoU threshold.
  // Next ID: 3
  message AveragePrecision {
    optional float iou_threshold = 1;
    optional float average_precision = 2;
  }

  // One entry for each in
  // ObjectDetectionAveragePrecisionParams::iou_thresholds, averaged over all
  // classes.
  repeated AveragePrecision individual_average_precisions = 1;
  // Average of Average Precision across all IoU thresholds.
  optional float overall_mean_average_precision = 2;
}

// Parameters that define how the Object Detection task is evaluated
// end-to-end.
//
// Next ID: 4
message ObjectDetectionParams {
  // Required.
  // Model's outputs should be same as a TFLite-compatible SSD model.
  // Refer:
  // https://www.tensorflow.org/lite/models/object_detection/overview#output
  // TODO(b/133772912): Generalize support for other types of object detection
  // models.
  optional TfliteInferenceParams inference_params = 1;
  // Optional. Used to match ground-truth categories with model output.
  // SSD Mobilenet V1 Model trained on COCO assumes class 0 is background class
  // in the label file and class labels start from 1 to number_of_classes+1.
  // Therefore, default value is set as 1.
  optional int32 class_offset = 2 [default = 1];
  optional ObjectDetectionAveragePrecisionParams ap_params = 3;
}

// Metrics from evaluation of the object detection task.
//
// Next ID: 5
message ObjectDetectionMetrics {
  optional LatencyMetrics pre_processing_latency = 1;
  optional LatencyMetrics inference_latency = 2;
  optional TfliteInferenceMetrics inference_metrics = 3;
  optional ObjectDetectionAveragePrecisionMetrics average_precision_metrics = 4;
}