Add Sigmoid Cross Entropy Loss

Sources/DeepLearning/Loss.swift

@@ -16,14 +16,38 @@
 import TensorFlow
 #endif
 
+/// Computes the mean squared error between logits and labels.
+///
+/// - Parameters:
+///   - logits: One-hot encoded outputs from a neural network.
+///   - labels: One-hot encoded values that correspond to the correct output.
 @differentiable
 public func meanSquaredError<Scalar: TensorFlowFloatingPoint>(
     predicted: Tensor<Scalar>, expected: Tensor<Scalar>) -> Tensor<Scalar> {
     return (expected - predicted).squared().mean()
 }
 
+/// Computes the softmax cross entropy (categorical cross entropy) between logits and labels.
+///
+/// - Parameters:
+///   - logits: One-hot encoded outputs from a neural network.
+///   - labels: One-hot encoded values that correspond to the correct output.
 @differentiable
 public func softmaxCrossEntropy<Scalar: TensorFlowFloatingPoint>(
     logits: Tensor<Scalar>, labels: Tensor<Scalar>) -> Tensor<Scalar> {
     return -(labels * logSoftmax(logits)).mean(alongAxes: 0).sum()
 }
+
+/// Computes the sigmoid cross entropy (binary cross entropy) between logits and labels.
+///
+/// - Parameters:
+///   - logits: Single continuous values from `0` to `1`.
+///   - labels: Integer values that correspond to the correct output.
+@differentiable
+public func sigmoidCrossEntropy<Scalar: TensorFlowFloatingPoint>(
+    logits: Tensor<Scalar>, labels: Tensor<Scalar>
+) -> Tensor<Scalar> {
+    let loss = labels * log(logits) +
+        (Tensor<Scalar>(1) - labels) * log(Tensor<Scalar>(1) - logits)
+    return -loss.mean(alongAxes: 0).sum()
+}