Updated models to use revised name for callable (SE-0253), fixed a few issues caused by swift-apis changes.

Autoencoder/main.swift

@@ -94,7 +94,7 @@ struct Autoencoder: Layer {
         activation: tanh)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let encoder = input.sequenced(through: encoder1, encoder2, encoder3, encoder4)
         return encoder.sequenced(through: decoder1, decoder2, decoder3, decoder4)
     }

CIFAR/Data.swift

@@ -29,9 +29,32 @@ func downloadCIFAR10IfNotPresent(to directory: String = ".") {
     }
 }
 
+extension Tensor where Scalar : _TensorFlowDataTypeCompatible {
+    public var _tfeTensorHandle: _AnyTensorHandle {
+        TFETensorHandle(_owning: handle._cTensorHandle)
+    }
+}
+
 struct Example: TensorGroup {
     var label: Tensor<Int32>
     var data: Tensor<Float>
+
+    init(label: Tensor<Int32>, data: Tensor<Float>) {
+        self.label = label
+        self.data = data
+    }
+
+    public init<C: RandomAccessCollection>(
+        _handles: C
+    ) where C.Element: _AnyTensorHandle {
+        precondition(_handles.count == 2)
+        let labelIndex = _handles.startIndex
+        let dataIndex = _handles.index(labelIndex, offsetBy: 1)
+        label = Tensor<Int32>(handle: TensorHandle<Int32>(handle: _handles[labelIndex]))
+        data = Tensor<Float>(handle: TensorHandle<Float>(handle: _handles[dataIndex]))
+    }
+
+    public var _tensorHandles: [_AnyTensorHandle] { [label._tfeTensorHandle, data._tfeTensorHandle] }
 }
 
 // Each CIFAR data file is provided as a Python pickle of NumPy arrays

CIFAR/Models.swift

@@ -29,7 +29,7 @@ struct PyTorchModel: Layer {
     var dense3 = Dense<Float>(inputSize: 84, outputSize: 10, activation: identity)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let convolved = input.sequenced(through: conv1, pool1, conv2, pool2)
         return convolved.sequenced(through: flatten, dense1, dense2, dense3)
     }
@@ -54,7 +54,7 @@ struct KerasModel: Layer {
     var dense2 = Dense<Float>(inputSize: 512, outputSize: 10, activation: identity)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let conv1 = input.sequenced(through: conv1a, conv1b, pool1, dropout1)
         let conv2 = conv1.sequenced(through: conv2a, conv2b, pool2, dropout2)
         return conv2.sequenced(through: flatten, dense1, dropout3, dense2)

CIFAR/ResNet.swift

@@ -37,7 +37,7 @@ struct Conv2DBatchNorm: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: conv, norm)
     }
 }
@@ -68,7 +68,7 @@ struct BasicBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let blocksReduced = blocks.differentiableReduce(input) { last, layer in
             relu(layer(last))
         }
@@ -97,7 +97,7 @@ struct ResNet: Layer {
     var classifier = Dense<Float>(inputSize: 64, outputSize: 10, activation: softmax)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let tmp = relu(inputLayer(input))
         let convolved = tmp.sequenced(through: basicBlock1, basicBlock2, basicBlock3)
         return convolved.sequenced(through: averagePool, flatten, classifier)

CIFAR/WideResNet.swift

@@ -41,7 +41,7 @@ struct BatchNormConv2DBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let firstLayer = conv1(relu(norm1(input)))
         return conv2(relu(norm2(firstLayer)))
     }
@@ -87,7 +87,7 @@ struct WideResNetBasicBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let blocksReduced = blocks.differentiableReduce(input) { last, layer in
             relu(layer(last))
         }
@@ -126,7 +126,7 @@ struct WideResNet: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = input.sequenced(through: l1, l2, l3, l4)
         let finalNorm = relu(norm(inputLayer))
         return finalNorm.sequenced(through: avgPool, flatten, classifier)

Catch/README.md

@@ -23,5 +23,5 @@ installed. Make sure you've added the correct version of `swift` to your path.
 To train the model, run:
 
 ```
-swift -O catch.swift
+swift -O main.swift
 ```

Catch/main.swift

@@ -53,7 +53,7 @@ struct Model: Layer {
                               generator: &rng)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: layer1, layer2)
     }
 }
@@ -83,7 +83,7 @@ extension CatchAgent {
         let (ŷ, backprop) = model.appliedForBackpropagation(to: x)
         let maxIndex = ŷ.argmax().scalarized()
 
-        let 𝛁loss = -log(Tensor(ŷ.max())).broadcast(like: ŷ) * previousReward
+        let 𝛁loss = -log(Tensor<Float>(ŷ.max())).broadcasted(like: ŷ) * previousReward
         let (𝛁model, _) = backprop(𝛁loss)
         optimizer.update(&model.allDifferentiableVariables, along: 𝛁model)
 

Gym/CartPole/main.swift

@@ -48,7 +48,7 @@ struct Net: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: l1, l2)
     }
 }

MNIST/main.swift

@@ -64,7 +64,7 @@ struct Classifier: Layer {
     var layer1b = Dense<Float>(inputSize: 128, outputSize: 10, activation: softmax)
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let convolved = input.sequenced(through: conv1a, conv1b, pool1)
         return convolved.sequenced(through: dropout1a, flatten, layer1a, dropout1b, layer1b)
     }

MiniGo/Models/GoModel.swift

@@ -61,7 +61,7 @@ struct ConvBN: Layer {
     }
 
     @differentiable
-    func call(_ input: Tensor<Float>) -> Tensor<Float> {
+    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
         return norm(conv(input))
     }
 }
@@ -90,7 +90,7 @@ struct ResidualIdentityBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Tensor<Float>) -> Tensor<Float> {
+    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
         var tmp = relu(layer1(input))
         tmp = layer2(tmp)
         return relu(tmp + input)
@@ -158,7 +158,7 @@ public struct GoModel: Layer {
     }
 
     @differentiable(wrt: (self, input), vjp: _vjpCall)
-    public func call(_ input: Tensor<Float>) -> GoModelOutput {
+    public func callAsFunction(_ input: Tensor<Float>) -> GoModelOutput {
         let batchSize = input.shape[0]
         var output = relu(initialConv(input))
 

ResNet/Data.swift

@@ -32,9 +32,32 @@ func downloadCIFAR10IfNotPresent(to directory: String = ".") {
     }
 }
 
+extension Tensor where Scalar : _TensorFlowDataTypeCompatible {
+    public var _tfeTensorHandle: _AnyTensorHandle {
+        TFETensorHandle(_owning: handle._cTensorHandle)
+    }
+}
+
 struct Example: TensorGroup {
     var label: Tensor<Int32>
     var data: Tensor<Float>
+
+    init(label: Tensor<Int32>, data: Tensor<Float>) {
+        self.label = label
+        self.data = data
+    }
+
+    public init<C: RandomAccessCollection>(
+        _handles: C
+    ) where C.Element: _AnyTensorHandle {
+        precondition(_handles.count == 2)
+        let labelIndex = _handles.startIndex
+        let dataIndex = _handles.index(labelIndex, offsetBy: 1)
+        label = Tensor<Int32>(handle: TensorHandle<Int32>(handle: _handles[labelIndex]))
+        data = Tensor<Float>(handle: TensorHandle<Float>(handle: _handles[dataIndex]))
+    }
+
+    public var _tensorHandles: [_AnyTensorHandle] { [label._tfeTensorHandle, data._tfeTensorHandle] }
 }
 
 // Each CIFAR data file is provided as a Python pickle of NumPy arrays

ResNet/ResNet50.swift

@@ -37,7 +37,7 @@ struct ConvBN: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: conv, norm)
     }
 }
@@ -65,7 +65,7 @@ struct ResidualBasicBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return layer2(relu(layer1(input)))
     }
 }
@@ -94,7 +94,7 @@ struct ResidualBasicBlockShortcut: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return layer2(relu(layer1(input))) + shortcut(input)
     }
 }
@@ -127,7 +127,7 @@ struct ResidualConvBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let tmp = relu(layer2(relu(layer1(input))))
         return relu(layer3(tmp) + shortcut(input))
     }
@@ -150,7 +150,7 @@ struct ResidualIdentityBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let tmp = relu(layer2(relu(layer1(input))))
         return relu(layer3(tmp) + input)
     }
@@ -175,7 +175,7 @@ struct ResidualIdentityBlockStack: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: block1, block2, block3, block4, block5)
     }
 }
@@ -218,7 +218,7 @@ struct ResNet18: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = maxPool(relu(l1(input)))
         let level2 = inputLayer.sequenced(through: l2a, l2b)
         let level3 = level2.sequenced(through: l3a, l3b)
@@ -274,7 +274,7 @@ struct ResNet34: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = maxPool(relu(l1(input)))
         let level2 = inputLayer.sequenced(through: l2a, l2b, l2c)
         let level3 = level2.sequenced(through: l3a, l3b, l3c, l3d)
@@ -326,7 +326,7 @@ struct ResNet50: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = maxPool(relu(l1(input)))
         let level2 = inputLayer.sequenced(through: l2a, l2b, l2c)
         let level3 = level2.sequenced(through: l3a, l3b, l3c, l3d)
@@ -383,7 +383,7 @@ struct ResNet101: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = maxPool(relu(l1(input)))
         let level2 = inputLayer.sequenced(through: l2a, l2b, l2c)
         let level3 = level2.sequenced(through: l3a, l3b, l3c, l3d)
@@ -441,7 +441,7 @@ struct ResNet152: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = maxPool(relu(l1(input)))
         let level2 = inputLayer.sequenced(through: l2a, l2b, l2c)
         let level3 = level2.sequenced(through: l3a, l3b, l3c, l3d)

ResNet/ResNetV2.swift

@@ -38,7 +38,7 @@ struct Conv2DBatchNorm: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: conv, norm)
     }
 }
@@ -60,7 +60,7 @@ struct BatchNormConv2D: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return conv(relu(norm(input)))
     }
 }
@@ -88,7 +88,7 @@ struct PreActivatedResidualBasicBlock: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: layer1, layer2)
     }
 }
@@ -117,7 +117,7 @@ struct PreActivatedResidualBasicBlockShortcut: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         return input.sequenced(through: layer1, layer2) + shortcut(input)
     }
 }
@@ -162,7 +162,7 @@ struct PreActivatedResNet18: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = input.sequenced(through: l1, maxPool)
         let level2 = inputLayer.sequenced(through: l2a, l2b)
         let level3 = level2.sequenced(through: l3a, l3b)
@@ -221,7 +221,7 @@ struct PreActivatedResNet34: Layer {
     }
 
     @differentiable
-    func call(_ input: Input) -> Output {
+    func callAsFunction(_ input: Input) -> Output {
         let inputLayer = input.sequenced(through: l1, maxPool)
         let level2 = inputLayer.sequenced(through: l2a, l2b, l2c)
         let level3 = level2.sequenced(through: l3a, l3b, l3c, l3d)