Minor correction to optimizer & add exp lr scheuler

Author: pxl-th

src/nn/adam.jl

@@ -83,10 +83,9 @@ function _step!(opt::Adam, θ::T, ∇::T, i; dispose::Bool) where T <: AbstractA
         "instead: `$(size(θ))` vs `$(size(∇))`."))
 
     # Debiasing.
-    lr = opt.lr * √(1f0 - opt.β2^opt.current_step) / (1f0 - opt.β1^opt.current_step)
     adam_step_kernel!(get_backend(opt))(
         opt.μ[i], opt.ν[i], θ, ∇,
-        opt.lr, opt.β1, opt.β2, opt.ϵ; ndrange=length(θ))
+        opt.lr, opt.β1, opt.β2, opt.ϵ, opt.current_step; ndrange=length(θ))
 
     dispose && KA.unsafe_free!(∇)
 
@@ -95,16 +94,29 @@ end
 
 @kernel function adam_step_kernel!(
     μ, ν, Θ, @Const(∇), lr::Float32,
-    β1::Float32, β2::Float32, ϵ::Float32,
+    β1::Float32, β2::Float32, ϵ::Float32, step::UInt32,
 )
     i = @index(Global)
     @inbounds ∇ᵢ = ∇[i]
-    @inbounds ωᵢ = Θ[i]
+    ∇ᵢ² = ∇ᵢ^2
 
-    ∇ᵢ² = ∇ᵢ * ∇ᵢ
     @inbounds μᵢ = μ[i] = β1 * μ[i] + (1f0 - β1) * ∇ᵢ
     @inbounds νᵢ = ν[i] = β2 * ν[i] + (1f0 - β2) * ∇ᵢ²
 
-    @inbounds Θ[i] = ωᵢ - (lr * μᵢ) / (√νᵢ + ϵ)
+    # Debiasing.
+    μ̂ = μᵢ / (1f0 - β1^step)
+    ν̂ = νᵢ / (1f0 - β2^step)
+
+    @inbounds ωᵢ = Θ[i]
+    @inbounds Θ[i] = ωᵢ - lr * μ̂ / (√ν̂ + ϵ)
 end
 
+function exp_scheduler(lr_start::Float32, lr_end::Float32, steps::Int)
+    function _scheduler(step::Int)
+        (step < 0 || (lr_start ≈ 0f0 && lr_end ≈ 0f0)) && return 0f0
+
+        t = clamp(Float32(step / steps), 0f0, 1f0)
+        return exp(log(lr_start) * (1 - t) + log(lr_end) * t)
+    end
+    return _scheduler
+end