Ex. 12.5

[szcf-weiya]

[szcf-weiya]

(a)

ESL-CN/code/Ex.12.5/main.jl

Lines 8 to 32 in 3a6336b

	function plot_all_curves()
	# https://github.com/JuliaGraphics/Colors.jl/blob/master/src/names_data.jl
	# colors = ["blue", "red", "orange", "violet", "torquoise"]
	ycolors = fill("blue", (1, size(y, 1))) # use row vector!!
	for i = 1:size(y, 1)
	if y[i] == "aa"
	continue
	elseif y[i] == "ao"
	ycolors[i] = "red"
	elseif y[i] == "dcl"
	ycolors[i] = "orange"
	elseif y[i] == "iy"
	ycolors[i] = "violet"
	elseif y[i] == "sh"
	ycolors[i] = "turquoise"
	end
	end
	p0 = plot(X', linecolors = ycolors, legend = false, title = "mix")
	p1 = plot(X[y .== "aa", :]', lc = "blue", legend = false, title ="aa")
	p2 = plot(X[y .== "ao", :]', lc = "red", legend = false, title = "ao")
	p3 = plot(X[y .== "dcl", :]', lc = "orange", legend = false, title = "dcl")
	p4 = plot(X[y .== "iy", :]', lc = "violet", legend = false, title = "iy")
	p5 = plot(X[y .== "sh", :]', lc = "turquoise", legend = false, title = "sh")
	return savefig(plot(p0, p1, p2, p3, p4, p5), "all_curves.png")
	end

[szcf-weiya]

(b)

[szcf-weiya]

(c)

I think the reason for putting all phonemes of a speaker into either train set or test set is to keep the test set independent from the train set, otherwise, the final accuracy tends to be higher since we have known the information of the speaker through some partial phonemes in the train set before performing testing on the test set.

ESL-CN/code/Ex.12.5/main.jl

Lines 131 to 156 in 3a6336b

	function evaluate()
	# divide into train and test set
	idx_train = Bool[ifelse(occursin("train", x), 1, 0) for x in data[:,end]]
	Xtrain, ytrain = X[idx_train,:], y[idx_train]
	Xtest, ytest = X[.~idx_train, :], y[.~idx_train]
	accs = zeros(3, 4)
	bestacc = 0.0
	bestB = nothing
	bestθ = nothing
	for (i, J) in enumerate([5, 10, 15])
	B = construct_B(256, J)
	for (j, K) in enumerate([1, 3, 5, 7])
	println("J = $J, K = $K")
	Cs, θs = kmeans(Xtrain, ytrain, B, K)
	ypred = classify(θs, Xtest, B)
	accs[i, j] = sum(ytest .== ypred) / length(ytest)
	println(freqtable(ytest, ypred))
	if bestacc < accs[i, j]
	bestacc = accs[i, j]
	bestB = B
	bestθ = θs
	end
	end
	end
	return accs, bestB, bestθ
	end

The accuracy is as follows:

julia> accs
3×4 Array{Float64,2}:
 0.811805  0.835757  0.851155  0.842601
 0.867408  0.875962  0.88195   0.890505
 0.867408  0.886228  0.885372  0.893926

and comparison between some contingency tables,

J = 5, K = 1
5×5 Named Array{Int64,2}
Dim1 ╲ Dim2 │  aa   ao  dcl   iy   sh
────────────┼────────────────────────
aa          │ 112   64    0    0    0
ao          │  68  186    5    4    0
dcl         │   0    4  185    6    0
iy          │  11   39    4  244   13
sh          │   0    0    0    2  222

J = 15, K = 7
5×5 Named Array{Int64,2}
Dim1 ╲ Dim2 │  aa   ao  dcl   iy   sh
────────────┼────────────────────────
aa          │ 130   46    0    0    0
ao          │  64  198    1    0    0
dcl         │   0    0  192    3    0
iy          │   1    0    8  301    1
sh          │   0    0    0    0  224

And I also plot the (smooth) prototypes,

which can be treated as the extracted features from the original data.