Skip to content

Commit

Permalink
Altered weighting vote calculations to be more true to the original 2…
Browse files Browse the repository at this point in the history 
…001 paper (hence removing logs and sqrts)
  • Loading branch information
@jakewilliami
jakewilliami committed Aug 25, 2020
1 parent f1f5b5e commit ac61810
Show file tree
Hide file tree
Showing 2 changed files with 42 additions and 21 deletions.
59 changes: 40 additions & 19 deletions src/Adaboost.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,12 +28,14 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
return `classifiers`: List of selected features [type: HaarLikeFeature]
=#

# get number of positive and negative images (and create a global variable of the total number of images——global for the @everywhere scope)
numPos = length(positiveIIs)
numNeg = length(negativeIIs)
global numImgs = numPos + numNeg

# get image height and width
imgHeight, imgWidth = size(positiveIIs[1])

# Maximum feature width and height default to image width and height
# Maximum feature width and height default to image width and height
# max_feature_height = img_height if max_feature_height == -1 else max_feature_height
# max_feature_width = img_width if max_feature_width == -1 else max_feature_width
Expand All @@ -44,7 +46,7 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
maxFeatureWidth = imgWidth
end

# Create initial weights and labels
# Initialise weights $w_{1,i} = \frac{1}{2m}, \frac{1}{2l}$, for $y_i=0,1$ for negative and positive examples respectively
posWeights = deepfloat(ones(numPos)) / (2 * numPos)
negWeights = deepfloat(ones(numNeg)) / (2 * numNeg)
#=
Expand All @@ -59,6 +61,8 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
```
The series of `deep*` functions——though useful in general——were designed from awkward arrays of tuples of arrays, which came about from a translation error in this case.
=#

# Concatenate positive and negative weights into one `weights` array
weights = vcat(posWeights, negWeights)
# println(weights)
# weights = vcat(posWeights, negWeights)
Expand All @@ -68,8 +72,9 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
labels = vcat(ones(numPos), ones(numNeg) * -1)
# labels = vcat([ones(numPos), ones(numNeg) * -1])

global images = positiveIIs + negativeIIs
# global images = vcat(positiveIIs, negativeIIs)
# get list of images (global because of @everywhere scope)
# global images = positiveIIs + negativeIIs
global images = vcat(positiveIIs, negativeIIs)

# Create features for all sizes and locations
global features = _create_features(imgHeight, imgWidth, minFeatureWidth, maxFeatureWidth, minFeatureHeight, maxFeatureHeight)
Expand All @@ -82,19 +87,22 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
end

# println(typeof(numImgs));println(typeof(numFeatures))
# create an empty array (of zeroes) with dimensions (numImgs, numFeautures)
global votes = zeros((numImgs, numFeatures)) # necessarily different from `zero.((numImgs, numFeatures))`

# bar = progressbar.ProgressBar()
# @everywhere numImgs begin
# println(votes)

# show progress bar
@everywhere begin
n = numImgs
processes = length(numImgs)
processes = length(numImgs) # i.e., hypotheses
p = Progress(n, 1) # minimum update interval: 1 second
for i in processes # bar(range(num_imgs)):
for t in processes # bar(range(num_imgs)):
# votes[i, :] = np.array(list(Pool(processes=None).map(partial(_get_feature_vote, image=images[i]), features)))
# votes[i, :] = Array(map(partial(getVote, images[i]), features))
votes[i, :] = Array(map(feature -> getVote(feature, images[i]), features))
votes[t, :] = Array(map(feature -> getVote(feature, images[t]), features))
# votes[i, :] = [map(feature -> getVote(feature, images[i]), features)]
next!(p)
end
Expand All @@ -108,11 +116,11 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi

n = numClassifiers
p = Progress(n, 1) # minimum update interval: 1 second
for _ in processes
for t in processes
# println(typeof(length(featureIndices)))
classificationErrors = zeros(length(featureIndices))

# normalize weights
# normalize the weights $w_{t,i}\gets \frac{w_{t,i}}{\sum_{j=1}^n w_{t,j}}$
# weights *= 1. / np.sum(weights)
# weights *= float(1) / sum(weights)
# weights = float(weights) / sum(weights) # ensure weights do not get rounded by element-wise making them floats
Expand All @@ -121,21 +129,21 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
# weights *= 1. / sum(weights)
weights = deepdiv(deepfloat(weights), deepsum(weights))

# select best classifier based on the weighted error
for f in 1:length(featureIndices)
fIDX = featureIndices[f]
# For each feature j, train a classifier $h_j$ which is restricted to using a single feature. The error is evaluated with respect to $w_j,\varepsilon_j = \sum_i w_i\left|h_j\left(x_i\right)-y_i\right|$
for j in 1:length(featureIndices)
fIDX = featureIndices[j]
# classifier error is the sum of image weights where the classifier is right
# error = sum(map(lambda img_idx: weights[img_idx] if labels[img_idx] != votes[img_idx, f_idx] else 0, range(num_imgs)))
# error = sum(imgIDX -> (labels[imgIDX] ≠ votes[imgIDX, fIDX]) ? weights[imgIDX] : 0, 1:numImgs)
# error = sum(map(lambda img_idx: weights[img_idx] if labels[img_idx] != votes[img_idx, f_idx] else 0, range(num_imgs)))
error = deepsum(map(imgIDX -> labels[imgIDX] ≠ votes[imgIDX, fIDX] ? weights[imgIDX] : 0, 1:numImgs))
ε = deepsum(map(imgIDX -> labels[imgIDX] ≠ votes[imgIDX, fIDX] ? weights[imgIDX] : 0, 1:numImgs))
# lambda img_idx: weights[img_idx] if labels[img_idx] != votes[img_idx, f_idx] else 0
# imgIDX -> (labels[imgIDX] ≠ votes[imgIDX, fIDX]) ? weights[imgIDX] : 0

classificationErrors[f] = error
classificationErrors[j] = ε
end

# get the best feature; i.e. the feature with the smallest error
# choose the classifier $h_t$ with the lowest error $\varepsilon_t$
minErrorIDX = argmin(classificationErrors) # returns the index of the minimum in the array
bestError = classificationErrors[minErrorIDX]
bestFeatureIDX = featureIndices[minErrorIDX]
Expand All @@ -144,16 +152,29 @@ function learn(positiveIIs::AbstractArray, negativeIIs::AbstractArray, numClassi
# println(weights)
# set feature weight
bestFeature = features[bestFeatureIDX]
featureWeight = 0.5 * log((1 - bestError) / bestError)
bestFeature.weight = featureWeight
# featureWeight = 0.5 * log((1 - bestError) / bestError)
featureWeight = (1 - bestError) / bestError # β
bestFeature.weight = featureWeight # need to element-wise alter the struct `weight`; else we get `setfield! immutable struct of type HaarLikeFeature cannot be changed`

# classifiers = vcat(classifiers, bestFeature)
classifiers = push!(classifiers, bestFeature)

# update image weights
# update image weights $w_{t+1,i}=w_{t,i}\beta_{t}^{1-e_i}$
# weights = list(map(lambda img_idx: weights[img_idx] * np.sqrt((1-best_error)/best_error) if labels[img_idx] != votes[img_idx, best_feature_idx] else weights[img_idx] * np.sqrt(best_error/(1-best_error)), range(num_imgs)))
# weights = (imgIDX -> (labels[imgIDX] ≠ votes[imgIDX, bestFeatureIDX]) ? weights[imgIDX]*sqrt((1-bestError)/bestError) : weights[imgIDX]*sqrt(bestError/(1-bestError)), 1:numImgs)
weights = map(imgIDX -> (labels[imgIDX] ≠ votes[imgIDX, bestFeatureIDX]) ? weights[imgIDX]*sqrt((1-bestError)/bestError) : weights[imgIDX]*sqrt(bestError/(1-bestError)), 1:numImgs)

weights = map(imgIDX -> (labels[imgIDX] ≠ votes[imgIDX, bestFeatureIDX]) ? weights[imgIDX] * featureWeight : weights[imgIDX] * featureWeight, 1:numImgs)

# β = ε / (1 - ε)
#
# e = nothing
#
# if labels[imgIDX] ≠ votes[imgIDX, bestFeatureIDX]
# weights[imgIDX] * sqrt((1-bestError)/bestError)
# else
# weights[imgIDX]*sqrt(bestError/(1-bestError)
# end


# remove feature (a feature can't be selected twice)
# feature_indexes.remove(best_feature_idx)
Expand Down
4 changes: 2 additions & 2 deletions src/HaarLikeFeature.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,8 +35,8 @@ struct HaarLikeFeature <: HaarFeatureType#struct HaarLikeFeature{T} <: HaarObjec
height::Int64
threshold::Int64
polarity::Int64
weight::Number# number because it can be -Inf due to log(0)
# weight::Int64
# weight::Number# number because it can be -Inf due to log(0)
weight::Int64
# weight::Any

# constructor; equivalent of __init__ method within class # ::CartesianIndex
Expand Down

1 comment on commit ac61810

@jakewilliami
Copy link
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is actually not true to the original paper. See issue #3.

Please sign in to comment.