Revert runtests

NEWS.md

@@ -20,6 +20,7 @@
 * Added [helper functions](https://github.com/FluxML/Flux.jl/pull/873) `Flux.convfilter` and `Flux.depthwiseconvfilter` to construct weight arrays for convolutions outside of layer constructors so as to not have to depend on the default layers for custom implementations.
 * `dropout` function now has a mandatory [active](https://github.com/FluxML/Flux.jl/pull/1263)
 keyword argument. The `Dropout` struct *whose behavior is left unchanged) is the recommended choice for common usage.
+* Connectionist temporal classification loss added
 * and many more fixes and additions...
 
 # v0.10.1 - v0.10.4

docs/src/models/losses.md

@@ -38,4 +38,5 @@ Flux.Losses.hinge_loss
 Flux.Losses.squared_hinge_loss
 Flux.Losses.dice_coeff_loss
 Flux.Losses.tversky_loss
+Flux.Losses.ctc
 ```

src/losses/Losses.jl

@@ -16,11 +16,12 @@ export mse, mae, msle,
     tversky_loss,
     dice_coeff_loss,
     poisson_loss,
-    hinge_loss, siquared_hinge_loss,
+    hinge_loss, squared_hinge_loss,
     ctc
 
 include("utils.jl")
 include("functions.jl")
 include("ctc.jl")
+if CUDA.functional() include("ctc-gpu.jl") end
 
 end #module

src/losses/ctc.jl

@@ -113,22 +113,36 @@ function ctc_(ŷ, y)
 
   # Fill bottom-right corner so bounding errors can be avoided
   # by starting `u` at `U′-1`
+<<<<<<< HEAD
   β[T,U′] = typedZero
   β[T,U′-1] = typedZero
 
   # start at T-1 so that β(T, u) = log(0) for all u < U′ - 1
   for t=(T-1):-1:1
     for u=U′:-1:1
       if u > 2t || u > U′ + 1
+=======
+  β[T,U′] = 0.0
+  β[T,U′-1] = 0.0
+
+  for t=T:-1:1
+    for u=U′:-1:1
+      if t == T && u >= U′ - 1
+        β[t,u] = 0.0
+      elseif t == T && u < U′ - 1
+>>>>>>> ff6ddf5fae0e3705814bea8e318717a3c6bb0343
         continue
       end
+<<<<<<< HEAD
 
       idx = u+2
       idx -= z′[u] == blank || (idx < U′ && z′[u+2] == z′[u])
       idx = min(idx, U′)
 
       v = [β[t+1,i] + ŷ[z′[i], t+1] for i=u:idx]
       β[t, u] = logsum(v)
+=======
+>>>>>>> ff6ddf5fae0e3705814bea8e318717a3c6bb0343
     end
   end
 

test/ctc-gpu.jl

@@ -1,13 +1,13 @@
 using Test
 using Flux
-using Flux: ctc_
+using Flux.Losses: ctc, ctc_
 using Zygote: gradient
 using LinearAlgebra
-using CuArrays
+using CUDA
 using Statistics
 
 # Custom function to check numerical gradient of ctc loss,
-# based on `ngradient` in `Tracker.jl`
+# based on `ngradient` in `Zygote.jl`
 # 
 # Needs to check loss as defined at a particular time step
 # related to the change in x because slight deviations in
@@ -40,16 +40,16 @@ end
   y_cu = CuArray(y)
 
   g1 = gradient(ctc, x_cu, y_cu)[1]
-  g1 = g1 |> collect
+  g1 = collect(g1)
 
-  g2 = ctc_ngradient(x, y)[1]
+  g2 = ctc_ngradient(x_cu, y_cu)[1] |> collect
 
   @test all(isapprox.(g1, g2, rtol=1e-5, atol=1e-5))
 
   # test that GPU loss matches CPU implementation
 
-  l1 = Flux.ctc_(x_cu, y_cu)[1]
-  l2 = Flux.ctc_(x, y)[1]
+  l1 = Flux.Losses.ctc_(x_cu, y_cu)[1]
+  l2 = Flux.Losses.ctc_(Float32.(x), y)[1]
 
   @test all(isapprox.(l1, l2, rtol=1e-5, atol=1e-5))
 
@@ -64,5 +64,14 @@ end
   ghat = gradient(ctc, x_cu, y_cu)[1] |> collect
 
   @test all(isapprox.(g, ghat, rtol=1e-5, atol=1e-5))
+
+  x_cu = [-3. 12. 8. 15.; 4. 20. -2. 20.; 8. -33. 6. 5.] |> CuArray
+  y_cu = [1 1 0 0; 0 0 1 1; 0 0 0 0] |> CuArray
+  @test ctc(x_cu, y_cu) ≈ 8.02519869363453
+
+  g = [-2.29294774655333e-06 -0.999662657278862 1.75500863563993e-06 0.00669284889063; 0.017985914969696 0.999662657278861 -1.9907078755387e-06 -0.006693150917307; -0.01798362202195 -2.52019580677916e-20 2.35699239251042e-07 3.02026677058789e-07]
+
+ ghat = gradient(ctc, x_cu, y_cu)[1] |> collect
+ @test all(isapprox.(g, ghat, rtol=1e-5, atol=1e-5))
 
 end

test/ctc.jl

@@ -1,6 +1,6 @@
 using Test
 using Flux
-using Flux: ctc_
+using Flux.Losses: ctc, ctc_
 using Zygote: gradient
 using LinearAlgebra
 

test/runtests.jl

@@ -5,7 +5,7 @@ using Random, Statistics, LinearAlgebra
 using IterTools: ncycle
 
 Random.seed!(0)
-#=
+
 @testset "Utils" begin
   include("utils.jl")
 end
@@ -21,13 +21,13 @@ end
 @testset "Data" begin
   include("data.jl")
 end
-=#
+
 @testset "Losses" begin
   include("losses.jl")
   include("ctc.jl")
   if Flux.use_cuda[] include("ctc-gpu.jl") end
 end
-#=
+
 @testset "Layers" begin
   include("layers/basic.jl")
   include("layers/normalisation.jl")
@@ -50,4 +50,3 @@ end
     doctest(Flux)
   end
 end
-=#