Subject: Add R2Score metric. (#8169)

tfjs-layers/src/exports_metrics.ts

@@ -314,3 +314,22 @@ export function MSE(yTrue: Tensor, yPred: Tensor): Tensor {
 export function mse(yTrue: Tensor, yPred: Tensor): Tensor {
   return losses.meanSquaredError(yTrue, yPred);
 }
+
+/**
+ * Computes R2 score.
+ *
+ * ```js
+ * const yTrue = tf.tensor2d([[0, 1], [3, 4]]);
+ * const yPred = tf.tensor2d([[0, 1], [-3, -4]]);
+ * const r2Score = tf.metrics.r2Score(yTrue, yPred);
+ * r2Score.print();
+ * ```
+ * @param yTrue Truth Tensor.
+ * @param yPred Prediction Tensor.
+ * @return R2 score Tensor.
+ *
+ * @doc {heading: 'Metrics', namespace: 'metrics'}
+ */
+export function r2Score(yTrue: Tensor, yPred: Tensor): Tensor {
+  return metrics.r2Score(yTrue, yPred);
+}

tfjs-layers/src/metrics.ts

@@ -17,9 +17,7 @@ import {Tensor, tidy} from '@tensorflow/tfjs-core';
 
 import * as K from './backend/tfjs_backend';
 import {NotImplementedError, ValueError} from './errors';
-import {categoricalCrossentropy as categoricalCrossentropyLoss, cosineProximity, meanAbsoluteError, meanAbsolutePercentageError, meanSquaredError, sparseCategoricalCrossentropy as sparseCategoricalCrossentropyLoss} from './losses';
-import {binaryCrossentropy as lossBinaryCrossentropy} from './losses';
-import {lossesMap} from './losses';
+import {binaryCrossentropy as lossBinaryCrossentropy, categoricalCrossentropy as categoricalCrossentropyLoss, cosineProximity, lossesMap, meanAbsoluteError, meanAbsolutePercentageError, meanSquaredError, sparseCategoricalCrossentropy as sparseCategoricalCrossentropyLoss} from './losses';
 import {LossOrMetricFn} from './types';
 import * as util from './utils/generic_utils';
 
@@ -112,6 +110,14 @@ export function sparseTopKCategoricalAccuracy(
   throw new NotImplementedError();
 }
 
+export function r2Score(yTrue: Tensor, yPred: Tensor): Tensor {
+  return tidy(() => {
+    const sumSquaresResiduals = yTrue.sub(yPred).square().sum();
+    const sumSquares = yTrue.sub(yTrue.mean()).square().sum();
+    return tfc.scalar(1).sub(sumSquaresResiduals.div(sumSquares));
+  });
+}
+
 // Aliases.
 export const mse = meanSquaredError;
 export const MSE = meanSquaredError;

tfjs-layers/src/metrics_test.ts

@@ -16,7 +16,7 @@ import {scalar, Tensor, tensor, tensor1d, tensor2d} from '@tensorflow/tfjs-core'
 
 import {setEpsilon} from './backend/common';
 import * as tfl from './index';
-import {binaryAccuracy, categoricalAccuracy, get, getLossOrMetricName} from './metrics';
+import {binaryAccuracy, categoricalAccuracy, get, getLossOrMetricName, r2Score} from './metrics';
 import {LossOrMetricFn} from './types';
 import {describeMathCPUAndGPU, describeMathCPUAndWebGL2, expectTensorsClose} from './utils/test_utils';
 
@@ -283,6 +283,27 @@ describeMathCPUAndGPU('recall metric', () => {
   });
 });
 
+describeMathCPUAndGPU('r2Score', () => {
+  it('1D', () => {
+    const yTrue = tensor1d([3, -0.5, 2, 7, 4.2, 8.5, 1.3, 2.8, 6.7, 9.0]);
+    const yPred = tensor1d([2.5, 0.0, 2.1, 7.8, 4.0, 8.2, 1.4, 2.9, 6.5, 9.1]);
+    const score = r2Score(yTrue, yPred);
+    expectTensorsClose(score, scalar(0.985));
+  });
+  it('2D', () => {
+    const yTrue = tensor2d([
+      [3, 2.5], [-0.5, 3.2], [2, 1.9], [7, 5.1], [4.2, 3.8], [8.5, 7.4],
+      [1.3, 0.6], [2.8, 2.1], [6.7, 5.3], [9.0, 8.7]
+    ]);
+    const yPred = tensor2d([
+      [2.7, 2.3], [0.0, 3.1], [2.1, 1.8], [6.8, 5.0], [4.1, 3.7], [8.4, 7.2],
+      [1.4, 0.7], [2.9, 2.2], [6.6, 5.2], [9.2, 8.9]
+    ]);
+    const score = r2Score(yTrue, yPred);
+    expectTensorsClose(score, scalar(0.995));
+  });
+});
+
 describe('metrics.get', () => {
   it('valid name, not alias', () => {
     expect(get('binaryAccuracy') === get('categoricalAccuracy')).toEqual(false);