Remove Compat and src/compat.jl (#318)

REQUIRE

@@ -1,5 +1,4 @@
 julia 0.6
-Compat 0.25.2
 Formatting
 BinDeps
 JSON

src/MXNet.jl

@@ -24,8 +24,6 @@ import Base: round, ceil, floor, cos, sin, abs, sign, exp, sqrt, exp, log, norm,
 
 include("base.jl")
 
-include("compat.jl")
-
 include("context.jl")
 include("util.jl")
 

src/optimizers/adadelta.jl

@@ -78,14 +78,14 @@ function update(self :: AdaDelta, index :: Int, weight :: NDArray,
 
   # Update state.acc as in RMSProp
   @inplace state.acc .*= self.opts.rho
-  @inplace state.acc .+= (1 - self.opts.rho) * @compatmul(grad, grad)
+  @inplace state.acc .+= (1 - self.opts.rho) * grad .* grad
 
   # Compute update using the "old" state.delta_acc
-  update = @compatmul(grad, sqrt(state.delta_acc + self.opts.epsilon)) ./
+  update = grad .* sqrt(state.delta_acc + self.opts.epsilon) ./
     (sqrt(state.acc + self.opts.epsilon))
   @inplace weight .+= -lr * update
 
   # update state.delta_acc using update
   @inplace state.delta_acc .*= self.opts.rho
-  @inplace state.delta_acc .+= (1 - self.opts.rho) * @compatmul(update, update)
+  @inplace state.delta_acc .+= (1 - self.opts.rho) * update .* update
 end

src/optimizers/adagrad.jl

@@ -61,6 +61,6 @@ function update(self :: AdaGrad, index :: Int, weight :: NDArray,
   lr = get_learning_rate(self.opts.lr_scheduler, self.state)
   grad = normalized_gradient(self.opts, self.state, weight, grad)
 
-  @inplace state .+= @compatmul(grad, grad)
+  @inplace state .+= grad .* grad
   @inplace weight .+= -lr * grad ./ (sqrt(state + self.opts.epsilon))
 end

src/optimizers/adam.jl

@@ -60,8 +60,8 @@ function update(self :: ADAM, index :: Int, weight :: NDArray, grad :: NDArray,
   lr = state.current_lr
   grad = normalized_gradient(self.opts, self.state, weight, grad)
 
-  state.mt = self.opts.beta1 * state.mt + (1 - self.opts.beta1) * grad
-  state.vt = self.opts.beta2 * state.vt + (1 - self.opts.beta2) * @compatmul(grad, grad)
+  state.mt = self.opts.beta1 * state.mt + (1 - self.opts.beta1) .* grad
+  state.vt = self.opts.beta2 * state.vt + (1 - self.opts.beta2) .* grad .* grad
 
   at = sqrt(1.0 - state.beta2Power)/(1.0 - state.beta1Power)
 

src/optimizers/nadam.jl

@@ -91,7 +91,7 @@ function update(self :: Nadam, index :: Int, weight :: NDArray,
   mt = state.mt / (1.0 - momentum_next)
 
   @inplace state.nt .*= self.opts.beta2
-  @inplace state.nt .+= (1.0 - self.opts.beta2) * @compatmul(grad, grad)
+  @inplace state.nt .+= (1.0 - self.opts.beta2) .* grad .* grad
   nt = state.nt / (1.0 - state.beta2Power)
   state.beta2Power *= self.opts.beta2
 

src/optimizers/rmsprop.jl

@@ -65,7 +65,7 @@ function update(self :: RMSProp, index :: Int, weight :: NDArray,
   grad = normalized_gradient(self.opts, self.state, weight, grad)
 
   @inplace state .*= self.opts.rho
-  @inplace state .+= (1 - self.opts.rho) * @compatmul(grad, grad)
+  @inplace state .+= (1 - self.opts.rho) * grad .* grad
 
   @inplace weight .+= -lr * grad ./ (sqrt(state + self.opts.epsilon))
 end