From f9e31a020c6cf65425dc0b0415a241dd946bfd31 Mon Sep 17 00:00:00 2001
From: Adarsh Kumar <45385384+AdarshKumar712@users.noreply.github.com>
Date: Mon, 2 Mar 2020 13:25:23 +0530
Subject: [PATCH] Updated huber_loss with other minute changes

---
 src/layers/stateless.jl | 47 +++++++++++++++++++----------------------
 1 file changed, 22 insertions(+), 25 deletions(-)

diff --git a/src/layers/stateless.jl b/src/layers/stateless.jl
index 592e2fa1..01b26a8a 100644
--- a/src/layers/stateless.jl
+++ b/src/layers/stateless.jl
@@ -16,38 +16,31 @@ mse(ŷ, y) = sum((ŷ .- y).^2) * 1 // length(y)
 
 
 """
-    msle(ŷ, y;ϵ1=eps.(Float64.(ŷ)),ϵ2=eps.(Float64.(y)))
+    msle(ŷ, y; ϵ1=eps.(Float64.(ŷ)))
 
-Mean Squared Logarithmic Error. Returns the mean of the squared logarithmic errors `sum((log.(ŷ+ϵ1).-log.(y+ϵ2)).^2) * 1 / length(y)`.<br>
-The ϵ1 and ϵ2 terms provide numerical stability. This error penalizes an under-predicted estimate greater than an over-predicted estimate.
+Mean Squared Logarithmic Error. Returns the mean of the squared logarithmic errors `sum((log.(ŷ+ϵ1) .- log.(y+ϵ2)).^2) * 1 / length(y)`.<br>
+The `ϵ` term provides numerical stability. This error penalizes an under-predicted estimate greater than an over-predicted estimate.
 """
-msle(ŷ, y;ϵ1=eps.(ŷ),ϵ2=eps.(eltype(ŷ).(y))) = sum((log.(ŷ+ϵ1).-log.(y+ϵ2)).^2) * 1 // length(y)
+msle(ŷ, y; ϵ=eps.(ŷ)) = sum((log.(ŷ+ϵ).-log.(y+ϵ)).^2) * 1 // length(y)
 
 
 
 """
-    huber_loss(ŷ, y,delta=1.0)
+    huber_loss(ŷ, y; delta=1.0)
+
+Computes the mean of the Huber loss given the prediction `ŷ` and true values `y`. By default, delta is set to 1.0.
 
-Computes the mean of the Huber loss. By default, delta is set to 1.0.
                     | 0.5*|(ŷ-y)|,   for |ŷ-y|<=delta
       Hubber loss = |
                     | delta*(|ŷ-y| - 0.5*delta),  otherwise
 
 [`Huber Loss`](https://en.wikipedia.org/wiki/Huber_loss).
 """
-function huber_loss(ŷ, y,delta=1.0)
-  abs_error = abs.(ŷ.-y)
-  dtype= eltype(ŷ)
-  delta = dtype(delta)
-  hub_loss = dtype(0)
-  for i in 1:length(y)
-    if (abs_error[i]<=delta)
-      hub_loss+=abs_error[i]^2*dtype(0.5)
-    else
-      hub_loss+=delta*(abs_error[i]- dtype(0.5*delta))
-    end
-  end
-  hub_loss*1//length(y)
+function huber_loss(ŷ, y; delta = eltype(ŷ)(1))
+   abs_error = abs.(ŷ.-y)
+   temp = abs_error.<delta
+   x = eltype(ŷ)(0.5)
+   hub_loss = sum(((abs_error.^2).*temp).*x .+ delta*(abs_error.- x*delta).*(1 .-temp)) * 1 // length(y)
 end
 
 function _crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat, weight::Nothing)
@@ -167,6 +160,7 @@ poisson(ŷ, y) = sum(ŷ .- y .* log.(ŷ)) *1 // size(y,2)
     hinge(ŷ, y)
 
 Measures the loss given the prediction `ŷ` and true labels `y` (containing 1 or -1). 
+Returns `sum((max.(0,1 .-ŷ .* y))) *1 // size(y, 2)`
 
 [Hinge Loss](https://en.wikipedia.org/wiki/Hinge_loss)
 See also [`squared_hinge`](@ref).
@@ -176,35 +170,38 @@ hinge(ŷ, y) = sum(max.(0, 1 .-  ŷ .* y)) *1 // size(y,2)
 """
     squared_hinge(ŷ, y)
 
-Computes squared hinge loss given the prediction `ŷ` and true labels `y` (conatining 1 or -1)
+Computes squared hinge loss given the prediction `ŷ` and true labels `y` (conatining 1 or -1).
+Returns `sum((max.(0,1 .-ŷ .* y)).^2) *1 // size(y, 2)`
 
 See also [`hinge`](@ref).
 """
 squared_hinge(ŷ, y) = sum((max.(0,1 .-ŷ .* y)).^2) *1//size(y,2)
 
 """
-    dice_coeff_loss(y_pred,y_true,smooth = 1)
+    dice_coeff_loss(y_pred, y_true, smooth = 1)
 
-Loss function used in Image Segmentation. Calculates loss based on dice coefficient. Similar to F1_score
+Loss function used in Image Segmentation. Calculates loss based on dice coefficient. Similar to F1_score.
+    
     Dice_Coefficient(A,B) = 2 * sum( |A*B| + smooth) / (sum( A^2 ) + sum( B^2 )+ smooth)
     Dice_loss = 1 - Dice_Coefficient
 
 Ref: [V-Net: Fully Convolutional Neural Networks forVolumetric Medical Image Segmentation](https://arxiv.org/pdf/1606.04797v1.pdf)
 """
-function dice_coeff_loss(y_pred,y_true,smooth=eltype(y_pred)(1.0))
+function dice_coeff_loss(y_pred, y_true; smooth=eltype(y_pred)(1.0))
     intersection = sum(y_true.*y_pred)
     return 1 - (2*intersection + smooth)/(sum(y_true.^2) + sum(y_pred.^2)+smooth)
 end
 
 """
-    tversky_loss(y_pred,y_true,beta = 0.7)
+    tversky_loss(y_pred, y_true, beta = 0.7)
 
 Used with imbalanced data to give more weightage to False negatives. Larger β weigh recall higher than precision (by placing more emphasis on false negatives)
+    
     tversky_loss(ŷ,y,beta) = 1 - sum(|y.*ŷ| + 1) / (sum(y.*ŷ + beta*(1 .- y).*ŷ + (1 .- beta)*y.*(1 .- ŷ))+ 1)
 
 Ref: [Tversky loss function for image segmentation using 3D fully convolutional deep networks](https://arxiv.org/pdf/1706.05721.pdf)
 """
-function tversky_loss(y_pred,y_true,beta = eltype(y_pred)(0.7))
+function tversky_loss(y_pred, y_true; beta = eltype(y_pred)(0.7))
     intersection = sum(y_true.*y_pred)
     return 1 - (intersection+1)/(sum(y_true.*y_pred + beta*(1 .- y_true).* y_pred + (1-beta).*y_true.*(1 .- y_pred))+1)
 end