Convert Autoencoder and Catch to use Sequential (#203)

Partially fixes #202.

Author: Shashi456

Autoencoder/main.swift

@@ -23,34 +23,20 @@ let imageHeight = 28
 let imageWidth = 28
 
 let outputFolder = "./output/"
-
-/// An autoencoder.
-struct Autoencoder: Layer {
-    var encoder1 = Dense<Float>(
-        inputSize: imageHeight * imageWidth, outputSize: 128,
-        activation: relu)
-
-    var encoder2 = Dense<Float>(inputSize: 128, outputSize: 64, activation: relu)
-    var encoder3 = Dense<Float>(inputSize: 64, outputSize: 12, activation: relu)
-    var encoder4 = Dense<Float>(inputSize: 12, outputSize: 3, activation: relu)
-
-    var decoder1 = Dense<Float>(inputSize: 3, outputSize: 12, activation: relu)
-    var decoder2 = Dense<Float>(inputSize: 12, outputSize: 64, activation: relu)
-    var decoder3 = Dense<Float>(inputSize: 64, outputSize: 128, activation: relu)
-
-    var decoder4 = Dense<Float>(
-        inputSize: 128, outputSize: imageHeight * imageWidth,
-        activation: tanh)
-
-    @differentiable
-    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
-        let encoder = input.sequenced(through: encoder1, encoder2, encoder3, encoder4)
-        return encoder.sequenced(through: decoder1, decoder2, decoder3, decoder4)
-    }
-}
-
 let dataset = MNIST(batchSize: batchSize, flattening: true)
-var autoencoder = Autoencoder()
+// An autoencoder.
+var autoencoder = Sequential {
+    // The encoder.
+    Dense<Float>(inputSize: imageHeight * imageWidth, outputSize: 128, activation: relu)
+    Dense<Float>(inputSize: 128, outputSize: 64, activation: relu)
+    Dense<Float>(inputSize: 64, outputSize: 12, activation: relu)
+    Dense<Float>(inputSize: 12, outputSize: 3, activation: relu)
+    // The decoder.
+    Dense<Float>(inputSize: 3, outputSize: 12, activation: relu)
+    Dense<Float>(inputSize: 12, outputSize: 64, activation: relu)
+    Dense<Float>(inputSize: 64, outputSize: 128, activation: relu)
+    Dense<Float>(inputSize: 128, outputSize: imageHeight * imageWidth, activation: tanh)
+}
 let optimizer = RMSProp(for: autoencoder)
 
 // Training loop

Catch/main.swift

@@ -43,30 +43,20 @@ protocol Agent: AnyObject {
     func step(observation: Observation, reward: Reward) -> Action
 }
 
-struct Model: Layer {
-    typealias Input = Tensor<Float>
-    typealias Output = Tensor<Float>
-
-    var layer1 = Dense<Float>(inputSize: 3, outputSize: 50, activation: sigmoid,
-                              generator: &rng)
-    var layer2 = Dense<Float>(inputSize: 50, outputSize: 3, activation: sigmoid,
-                              generator: &rng)
-
-    @differentiable
-    func callAsFunction(_ input: Input) -> Output {
-        return input.sequenced(through: layer1, layer2)
-    }
-}
-
 class CatchAgent: Agent {
     typealias Action = CatchAction
 
-    var model: Model = Model()
-    let optimizer: Adam<Model>
+    var model = Sequential {
+        Dense<Float>(inputSize: 3, outputSize: 50, activation: sigmoid, generator: &rng)
+        Dense<Float>(inputSize: 50, outputSize: 3, activation: sigmoid, generator: &rng)
+    }
+
+    var learningRate: Float
+    lazy var optimizer = Adam(for: self.model, learningRate: self.learningRate)
     var previousReward: Reward
 
     init(initialReward: Reward, learningRate: Float) {
-        optimizer = Adam(for: model, learningRate: learningRate)
+        self.learningRate = learningRate
         previousReward = initialReward
     }
 }