Update min(_:_:) and max(_:_:) gradients to match Python TensorFlow (tensorflow#480)

Author: jon-tow

Sources/TensorFlow/Loss.swift

@@ -287,8 +287,6 @@ public func sigmoidCrossEntropy<Scalar: TensorFlowFloatingPoint>(
 ) -> Tensor<Scalar> {
     // This numerically stable implementation is based on the TensorFlow Python API.
     let maxLogitsWithZero = max(logits, Tensor(0))
-    // Note: `result` is split into two lines to avoid the "compiler is unable to type-check this
-    // expression in reasonable time" error.
-    let result = log(1 + exp(-abs(logits)))
-    return reduction(maxLogitsWithZero - logits * labels + result)
+    let negAbsLogits = max(logits, -logits) // Custom `abs` to compute gradients at `0`.
+    return reduction(maxLogitsWithZero - logits * labels + log1p(exp(-negAbsLogits)))
 }

Sources/TensorFlow/Operators/Math.swift

@@ -1344,7 +1344,9 @@ internal func _vjpMax<T: TensorFlowFloatingPoint>(
     _ y: Tensor<T>
 ) -> (Tensor<T>, (Tensor<T>) -> (Tensor<T>, Tensor<T>)) {
     let value = max(x, y)
-    return (value, { v in _vjpMinMaxHelper(x, y, originalValue: value, seed: v) })
+    return (value, { v in
+        _vjpMinMaxHelper(x, y, originalValue: value, seed: v, comparisonOperation: .>=)
+    })
 }
 
 /// Returns the element-wise maximum of the scalar and the tensor, broadcasting the scalar.
@@ -1375,7 +1377,9 @@ internal func _vjpMin<T: TensorFlowFloatingPoint>(
     _ y: Tensor<T>
 ) -> (Tensor<T>, (Tensor<T>) -> (Tensor<T>, Tensor<T>)) {
     let value = min(x, y)
-    return (value, { v in _vjpMinMaxHelper(x, y, originalValue: value, seed: v) })
+    return (value, { v in
+        _vjpMinMaxHelper(x, y, originalValue: value, seed: v, comparisonOperation: .<=)
+    })
 }
 
 /// Returns the element-wise minimum of the scalar and the tensor, broadcasting the scalar.
@@ -1397,11 +1401,12 @@ internal func _vjpMinMaxHelper<T: TensorFlowFloatingPoint>(
     _ x: Tensor<T>,
     _ y: Tensor<T>,
     originalValue: Tensor<T>,
-    seed: Tensor<T>
+    seed: Tensor<T>,
+    comparisonOperation: (Tensor<T>, Tensor<T>) -> Tensor<Bool>
 ) -> (Tensor<T>, Tensor<T>) {
-    let denominator = 1 + Tensor<T>(x .== y)
-    let lhsGrad = seed * Tensor<T>(x .== originalValue) / denominator
-    let rhsGrad = seed * Tensor<T>(y .== originalValue) / denominator
+    let mask = Tensor<T>(comparisonOperation(x, y))
+    let lhsGrad = seed * mask
+    let rhsGrad = seed * (1 - mask)
     let (lhsShape, rhsShape) = (x.shapeTensor, y.shapeTensor)
     let (lhsAxes, rhsAxes) = Raw.broadcastGradientArgs(s0: lhsShape, s1: rhsShape)
     return (lhsGrad.sum(squeezingAxes: lhsAxes).reshaped(toShape: lhsShape),

Tests/TensorFlowTests/LossTests.swift

@@ -196,27 +196,25 @@ final class LossTests: XCTestCase {
         assertEqual(loss, Tensor(expectedLoss), accuracy: 1e-6)
     }
 
-    func testSigmoidCrossEntropyGrad() {
-        let logits = Tensor<Float>(
-            shape: [2, 4],
-            scalars: [-100, -2, -2, 0, 2, 2, 2, 100])
-
-        let labels = Tensor<Float>(
-            shape: [2, 4],
-            scalars: [0, 0, 1, 0, 0, 1, 0.5, 1])
+    func testSigmoidCrossEntropyGradient() {
+        let logits = Tensor<Float>(shape: [2, 4], scalars: [-100, -2, -2, 0, 0, 2, 2, 100])
+        let labels = Tensor<Float>(shape: [2, 4], scalars: [0, 0, 1, 0, 1, 1, 0.5, 1])
 
-        // For each element x in logits and y in labels, the gradient is sigmoid(x) - y.
-        let expectedGradientsBeforeMean = Tensor<Float>(
-            shape: [2, 4],
-            scalars: [0.00,  0.11920291, -0.8807971,  0.5,
-                      0.8807971, -0.11920291,  0.3807971 , 0.0])
-
-        // As the loss is mean loss, we should scale the golden gradient numbers.
-        let expectedGradients = expectedGradientsBeforeMean / Float(logits.scalars.count)
-        let gradients = gradient(
+        let computedGradient = gradient(
             at: logits,
             in: { sigmoidCrossEntropy(logits: $0, labels: labels) })
-        assertEqual(gradients, expectedGradients, accuracy: 1e-6)
+        // The expected value of the gradient was computed using Python TensorFlow 1.14 with
+        // the following code:
+        // ```
+        // with tf.GradientTape() as t:
+        //    t.watch([logits])
+        //    y = tf.losses.sigmoid_cross_entropy(labels, logits, reduction="weighted_mean")
+        // print(t.gradient(y, [logits]))
+        // ```
+        let expectedGradient = Tensor<Float>([
+            [0.0, 0.01490036, -0.11009964, 0.0625],
+            [-0.0625, -0.01490036, 0.04759964, 0.0]])
+        assertEqual(computedGradient, expectedGradient, accuracy: 1e-6)
     }
 
     static var allTests = [
@@ -238,6 +236,6 @@ final class LossTests: XCTestCase {
         ("testSoftmaxCrossEntropyWithProbabilitiesGrad",
          testSoftmaxCrossEntropyWithProbabilitiesGrad),
         ("testSigmoidCrossEntropyLoss", testSigmoidCrossEntropyLoss),
-        ("testSigmoidCrossEntropyGrad", testSigmoidCrossEntropyGrad),
+        ("testSigmoidCrossEntropyGradient", testSigmoidCrossEntropyGradient),
     ]
 }

Tests/TensorFlowTests/TensorAutoDiffTests.swift

@@ -153,6 +153,77 @@ final class TensorAutoDiffTests: XCTestCase {
         XCTAssertEqual(varianceGradAlongAxes(input), expected)
     }
 
+    func testMin() {
+        // The expected gradient values were computed using the following TensorFlow 2.0 Beta1
+        // Python code with respective `a` and `b` tensors:
+        // ```
+        // with tf.GradientTape() as t:
+        //     t.watch([a, b])
+        //     y = tf.math.reduce_sum(tf.minimum(a, b))
+        // print(t.gradient(y, [a, b]))
+        // ```
+        do {
+            let a = Tensor<Float>([4, 5, 3])
+            let b = Tensor<Float>([4, 2, 6])
+            let computedGradient1 = gradient(at: a, b) { a, b in min(a, b).sum() }
+            let expectedGradient1: (Tensor<Float>, Tensor<Float>) = (
+                [1.0, 0.0, 1.0], [0.0, 1.0, 0.0])
+            XCTAssertEqual(computedGradient1.0, expectedGradient1.0)
+            XCTAssertEqual(computedGradient1.1, expectedGradient1.1)
+
+            let computedGradient2 = gradient(at: a, b) { a, b in min(b, a).sum() }
+            let expectedGradient2: (Tensor<Float>, Tensor<Float>) =  (
+                [0.0, 0.0, 1.0], [1.0, 1.0, 0.0])
+            XCTAssertEqual(computedGradient2.0, expectedGradient2.0)
+            XCTAssertEqual(computedGradient2.1, expectedGradient2.1)
+        }
+
+        do {
+            let a = Tensor<Float>([[3.0, -2.0], [0.3, 10.0]])
+            let b = Tensor<Float>([9.0, -3.0])
+            let computedGradient = gradient(at: a, b) { a, b in min(a, b).sum() }
+            let expectedGradient: (Tensor<Float>, Tensor<Float>) = (
+                [[1.0, 0.0], [1.0, 0.0]], [0.0, 2.0])
+            XCTAssertEqual(computedGradient.0, expectedGradient.0)
+            XCTAssertEqual(computedGradient.1, expectedGradient.1)
+        }
+    }
+
+    func testMax() {
+        // The expected gradient values were computed using the following TensorFlow 2.0 Beta1
+        // Python code with respective `a` and `b` tensors:
+        // ```
+        // with tf.GradientTape() as t:
+        //     t.watch([a, b])
+        //     y = tf.math.reduce_sum(tf.maximum(a, b))
+        // print(t.gradient(y, [a, b]))
+        // ```
+        do {
+            let a = Tensor<Float>([4, 5, 3])
+            let b = Tensor<Float>([4, 2, 6])
+            let computedGradient1 = gradient(at: a, b) { a, b in max(a, b).sum() }
+            let expectedGradient1: (Tensor<Float>, Tensor<Float>) = (
+                [1.0, 1.0, 0.0], [0.0, 0.0, 1.0])
+            XCTAssertEqual(computedGradient1.0, expectedGradient1.0)
+            XCTAssertEqual(computedGradient1.1, expectedGradient1.1)
+
+            let computedGradient2 = gradient(at: a, b) { a, b in max(b, a).sum() }
+            let expectedGradient2: (Tensor<Float>, Tensor<Float>) = (
+                [0.0, 1.0, 0.0],  [1.0, 0.0, 1.0])
+            XCTAssertEqual(computedGradient2.0, expectedGradient2.0)
+            XCTAssertEqual(computedGradient2.1, expectedGradient2.1)
+        }
+        do {
+            let a = Tensor<Float>([[3.0, -2.0], [0.3, 10.0]])
+            let b = Tensor<Float>([9.0, -3.0])
+            let computedGradient = gradient(at: a, b) { a, b in max(a, b).sum() }
+            let expectedGradient: (Tensor<Float>, Tensor<Float>)  = (
+                [[0.0, 1.0], [0.0, 1.0]], [2.0, 0.0])
+            XCTAssertEqual(computedGradient.0, expectedGradient.0)
+            XCTAssertEqual(computedGradient.1, expectedGradient.1)
+        }
+    }
+
     /*TODO:(https://bugs.swift.org/browse/TF-771): Disabling this case as assertions fail.
     func testTensorInitStacking() {
         let a1 = Tensor<Float>([1, 2, 3, 4, 5])
@@ -449,6 +520,8 @@ final class TensorAutoDiffTests: XCTestCase {
         ("testSum", testSum),
         ("testMean", testMean),
         ("testVariance", testVariance),
+        ("testMin", testMin),
+        ("testMax", testMax),
         // TODO(https://bugs.swift.org/browse/TF-771): Disabling the failing test.
         // ("testTensorInitStacking", testTensorInitStacking),
         ("testExpandingShape", testExpandingShape),