Merge pull request #11 from pedromxavier/px/upgrade-images

Upgrade Mental Images Interface

docs/src/examples.md

@@ -29,7 +29,7 @@ test_x = Array{Bool}[test_x[:, :, i] .>= 0.25 for i = 1:10_000]
 
 train!(wnn, train_y, train_x)
 
-s = count(classify(wnn, test_x; bleach=12) .== test_y)
+s = count(classify(wnn, test_x; bleach=10) .== test_y)
 
 println("Accuracy: $(100. * s / 10_000)%")
 ```
@@ -38,23 +38,17 @@ println("Accuracy: $(100. * s / 10_000)%")
 ```@example mnist
 using Images
 
-images = WiSARD.images(wnn)
+images = WiSARD.images(RGB, wnn; w=28, h=28)
 
-img = images[0]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[0]
 ```
 
 ```@example mnist
-img = images[1]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[1]
 ```
 
 ```@example mnist
-img = images[2]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[2]
 ```
 
 ## Fashion MNIST
@@ -82,7 +76,7 @@ test_x = Array{Bool}[test_x[:, :, i] .>= 0.25 for i = 1:10_000]
 
 train!(wnn, train_y, train_x)
 
-s = count(classify(wnn, test_x; bleach=12) .== test_y)
+s = count(classify(wnn, test_x; bleach=10) .== test_y)
 
 println("Accuracy: $(100. * s / 10_000)%")
 ```
@@ -91,21 +85,15 @@ println("Accuracy: $(100. * s / 10_000)%")
 ```@example fashion-mnist
 using Images
 
-images = WiSARD.images(wnn)
+images = WiSARD.images(RGB, wnn; w=28, h=28)
 
-img = images[0]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[0]
 ```
 
 ```@example fashion-mnist
-img = images[1]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[1]
 ```
 
 ```@example fashion-mnist
-img = images[2]
-
-RGB.([img[(i - 1) * 28 + j] for i = 1:28, j = 1:28])
+images[2]
 ```

src/images/images.jl

@@ -2,7 +2,21 @@ function address(wnn::WNN{S, T}, k::T) where {S, T}
     [j for j = 1:wnn.d if (k >> (j - 1)) % 2 == 1]
 end
 
-function images(wnn::WNN{S, T}, y::S) where {S, T}
+function images(
+    N::Type{<:Any},
+    wnn::WNN{S, T},
+    y::S;
+    w::Union{Int, Nothing} = nothing,
+    h::Union{Int, Nothing} = nothing,
+    ) where {S, T}
+    w, h = if isnothing(w) && isnothing(h)
+        1, wnn.n * wnn.d
+    elseif isnothing(w) || isnothing(h)
+        error("Both width 'w' and height 'h' must be informed, or none of them")
+    else
+        w, h
+    end
+
     cls = wnn.cls[y]
     img = zeros(Int, wnn.n * wnn.d)
 
@@ -17,12 +31,34 @@ function images(wnn::WNN{S, T}, y::S) where {S, T}
     end
 
     return if M == 0
-        Float64.(img)
+        N[img[(i - 1) * w + j] for i=1:h, j=1:w]
     else
-        Float64.(img ./ M)
+        N[img[(i - 1) * w + j] / M for i=1:h, j=1:w]
     end
 end
 
-function images(wnn::WNN{S, T}) where {S, T}
-    return Dict{S, Array{Float64}}(y => images(wnn, y) for y in keys(wnn.cls))
+function images(
+    wnn::WNN{S, T},
+    y::S;
+    w::Union{Int, Nothing} = nothing,
+    h::Union{Int, Nothing} = nothing,
+    ) where{S, T}
+    images(Float64, wnn, y; w=w, h=h)
+end
+
+function images(
+    N::Type{<:Any},
+    wnn::WNN{S, T};
+    w::Union{Int, Nothing} = nothing,
+    h::Union{Int, Nothing} = nothing,
+    ) where {S, T}
+    Dict{S, Array{N}}(y => images(N, wnn, y; w=w, h=h) for y in keys(wnn.cls))
+end
+
+function images(
+    wnn::WNN;
+    w::Union{Int, Nothing} = nothing,
+    h::Union{Int, Nothing} = nothing,
+    )
+    images(Float64, wnn; w=w, h=h)
 end

src/model/model.jl

@@ -45,9 +45,6 @@ end
 Base.show(io::IO, wnn::WNN{S, T}) where {S <: Any, T <: BigInt}  = print(io, "WNN[∞ bits, $(wnn.d) × $(wnn.n)]")
 Base.show(io::IO, wnn::WNN{S, T}) where {S <: Any, T <: Unsigned} = print(io, "WNN[$(T.size * 8) bits, $(wnn.d) × $(wnn.n)]")
 
-@doc raw"""
-    address(wnn::WNN, x::AbstractArray, i::Int)
-"""
 function address(wnn::WNN, x::AbstractArray, i::Int)
     @inbounds sum(iszero(x[wnn.map[(i - 1) * wnn.d + j]]) ? 0 : 1 << (j - 1) for j = 1:wnn.d)
 end