Change momentum and epsilon properties to scalars. (#220)

Author: dan-zheng

MiniGo/Models/GoModel.swift

@@ -54,10 +54,7 @@ struct ConvBN: Layer {
         // TODO(jekbradbury): thread through bias and affine boolean arguments
         // (behavior is correct for inference but this should be changed for training)
         self.conv = Conv2D(filterShape: filterShape, strides: strides, padding: padding)
-        self.norm = BatchNorm(
-            featureCount: filterShape.3,
-            momentum: Tensor<Float>(0.95),
-            epsilon: Tensor<Float>(1e-5))
+        self.norm = BatchNorm(featureCount: filterShape.3, momentum: 0.95, epsilon: 1e-5)
     }
 
     @differentiable

Transformer/Model.swift

@@ -227,10 +227,10 @@ struct EncoderLayer: Layer {
             size: size,
             headCount: headCount)
         selfAttentionDropout = Dropout(probability: dropProbability)
-        selfAttentionNorm = LayerNorm(featureCount: size, axis: 2, epsilon: Tensor<Float>(1e-5))
+        selfAttentionNorm = LayerNorm(featureCount: size, axis: 2, epsilon: 1e-5)
         feedForward = FeedForward(size: size, hidden: 4 * size, dropProbability: dropProbability)
         feedForwardDropout = Dropout(probability: dropProbability)
-        feedForwardNorm = LayerNorm(featureCount: size, axis: 2, epsilon: Tensor<Float>(1e-5))
+        feedForwardNorm = LayerNorm(featureCount: size, axis: 2, epsilon: 1e-5)
     }
 
     @differentiable(wrt: (self, input))

Transformer/PythonCheckpointReader.swift

@@ -73,7 +73,7 @@ extension LayerNorm: InitializableFromPythonCheckpoint {
             offset: readTensor(fromPath: path, name: scope + "/b", scalarType: Scalar.self),
             scale: readTensor(fromPath: path, name: scope + "/g", scalarType: Scalar.self),
             axis: -1,
-            epsilon: Tensor(1e-5))
+            epsilon: 1e-5)
     }
 }