def testClusters(self, x):
p = self.model.predict(x)
print 'Number of clusters: ', len(self.clusters)
for i in xrange(len(self.clusters)):
print core.clusterIntersectionRatio(x, self.clusters[i])
C = self.clusters[p]
print 'This should be better:'
print core.clusterIntersectionRatio(x, C)
print 'This should be even better:'
D = self.buildCluster(x)
print core.clusterIntersectionRatio(x, D)
def test(data, recommender, fractionTrain=.8, highFactor=.1, verbose=False):
"""
Parameters:
cluster: an array of arrays
fractionTrain: a float specifying percent of data to use as train
highFactor: ratio of any given element to the largest element in the array
Returns:
the RMSE produced by removing a specific color, then adding it back
Algorithm:
Works by first finding the max value in the histogram and then trying to find the
index of the histogram that contains the value that is closest (in terms of a ration
with the max value) to highFactor. Removes this color, then passes it to recommender
to get back a value which it then adds to the histogramogram, then takes the rmse between
the original and the modified
"""
xTrain, xTest = ml.splitData(data, fractionTrain)
n = xTest.shape[0]
m = xTrain.shape[0]
train_colors, _, train_histograms = removeColors(xTrain, highFactor=highFactor)
recommender.fit(train_histograms, train_colors)
if verbose:
print 'Done fitting'
colors, quantities, histograms = removeColors(xTest, highFactor=highFactor)
assert(colors.shape[0] == n)
assert(histograms.shape[0] == n)
numCorrect = 0
D = histograms.shape[1]
count = np.zeros(D)
for color in colors:
count[color] += 1
tmp = count[np.where(count > 0)]
color_mean = np.mean(tmp)
color_stdev = np.std(tmp)
print 'Color mean and stdev', color_mean, color_stdev
recommendedColors = np.zeros((n))
ignored = 0
intersectionRatio = 0.
for i in xrange(n):
if i % 100 == 1:
print 'Partial %d: %f' % (i, float(numCorrect) / (i - ignored + 1))
color, amount = colors[i], quantities[i]
# Ignore colors that might bias us
if count[color] > color_mean + color_stdev:
ignored += 1
continue
hist = histograms[i]
# Ignore colors that are basically the background
if hist[color] > 0.4:
ignored += 1
continue
if verbose:
print 'Testing site %s' % names[i]
print 'Amount remmoved %d' % amount
try:
cluster = recommender.cluster(hist)
intersectionRatio += core.clusterIntersectionRatio(hist, cluster)
except:
pass
#recommender.testClusters(hist)
recommendedColor = recommender.predict(hist)
r1, g1, b1 = image.binToRGB(color)
r2, g2, b2 = image.binToRGB(recommendedColor)
if verbose:
print 'Removed color %d %d %d. Recommended color %d %d %d.' % (r1, g1, b1, r2, g2, b2)
print 'Color distance: %d' % (image.binDistance(recommendedColor, color))
recommendedColors[i] = recommendedColor
if verbose:
print 'Recommended color %d' % (recommendedColor)
if recommendedColor == color:
numCorrect += 1
print 'Ignored: %d. Used: %d' % (ignored, n - ignored)
print 'Mean cluster intersection ratio: %f' % (intersectionRatio / (n - ignored))
print colorError(colors, recommendedColors)
percentCorrect = float(numCorrect)/(n - ignored)
return percentCorrect