def findKMeansPerLabel(activations, k, labelCount, targetPath, indexLabelMapping):
labels = None
if indexLabelMapping != None:
labels = utility.getLabelStrings(indexLabelMapping)
# split activations by label
activationsByLabel = []
counter = 0
while counter < labelCount:
currentLabelIndex = activations.shape[1] - labelCount + counter
logger.debug(currentLabelIndex)
currentSelection = activations[activations[:, currentLabelIndex] == 1]
activationsByLabel.append(currentSelection)
counter += 1
counter = 0
clusters = []
iterations = []
for batch in activationsByLabel:
if labels != None:
logger.info('Running KMeans for label ' + labels[counter] + '.')
else:
logger.info('Running KMeans for label ' + str(counter))
logger.debug("Batch shape: " + str(batch.shape))
[c, i] = kMeans_core.runKMeans(batch, labelCount, k, MAX_ITERATIONS)
clusters.extend(c)
iterations.append(i)
counter += 1
kMeans_core.saveResults(clusters, iterations, targetPath)
return [clusters, iterations]
def runTests(activationsPath, indexLabelMappingPath, sourcePath):
labels = utility.getLabelStrings(indexLabelMappingPath)
activations = utility.arrayFromFile(activationsPath)
if sourcePath.endswith('/') == False:
sourcePath += '/'
clusterGroups = loadClusters(sourcePath) # returns list with all previously generated cluster groups
mixedClusterResultsSimple = []
perLabelClusterResultsSimple = []
for clusters in clusterGroups:
logger.info("Evaluating clusters at " + clusters['path'])
if(clusters['type'] == 'perLabel'):
[confusionMatrix, meanAveragePrecision, overallCorrect, overallWrong] = runTest(clusters['data'], activations, labels True) # (clusters, activations, labels)
perLabelClusterResultsSimple.append([clusters['k'], meanAveragePrecision, overallCorrect, overallWrong])
saveConfusionMatrix(confusionMatrix, clusters['path'] + "confusion_test.npy")
utility.plotConfusionMatrix(confusionMatrix, labels, clusters['path'] + "confusion_test.pdf")
else:
[confusionMatrix, meanAveragePrecision, overallCorrect, overallWrong] = runTest(clusters['data'], activations, labels False) # (clusters, activations, labels)
mixedClusterResultsSimple.append([clusters['k'], meanAveragePrecision, overallCorrect, overallWrong])
saveConfusionMatrix(confusionMatrix, clusters['path'] + "confusion_test.npy")
utility.plotConfusionMatrix(confusionMatrix, labels, clusters['path'] + "confusion_test.pdf")
overviewPerLabel = np.array(perLabelClusterResultsSimple)
overviewMixed = np.array(mixedClusterResultsSimple)
saveOverview(overviewPerLabel, sourcePath + 'perLabel_')
saveOverview(overviewMixed, sourcePath + 'result_mixed_')
utility.plotKMeansOverview(overviewPerLabel, sourcePath + 'perLabel_result.pdf', True)
utility.plotKMeansOverview(overviewMixed, sourcePath + 'result_mixed_result.pdf', True)
def plotOverview(activations, labelCount, indexLabelMappingPath, plotFileName):
firstLabelIndex = activations.shape[1] - labelCount
reduceActivationsBy = 1
if firstLabelIndex % 4 == 0 and firstLabelIndex > 1024:
reduceActivationsBy = 4
elif firstLabelIndex % 2 == 0 and firstLabelIndex > 1024:
reduceActivationsBy = 2
reduceActivationsTo = firstLabelIndex / reduceActivationsBy
activationsPerLabel = 1024 / labelCount
logger.info(str(activationsPerLabel) + " rows per label.")
labels = np.array(activations)[:,firstLabelIndex:]
selection = np.empty((0, activations.shape[1]))
tickLabels = []
labelCounter = 0
while labelCounter < labels.shape[1]:
picked = np.random.randint(0,activations.shape[0],2)
monoLabelSelection = activations[np.logical_or.reduce([activations[:,firstLabelIndex+labelCounter] == 1])]
picked = np.random.randint(0, monoLabelSelection.shape[0], activationsPerLabel)
subSelection = monoLabelSelection[picked]
selection = np.vstack((selection, subSelection))
tickLabels.append("Label " + str(labelCounter))
labelCounter += 1
if indexLabelMappingPath != None:
tickLabels = utility.getLabelStrings(indexLabelMappingPath)
cmap = plt.get_cmap('Greys_r')
cmap_adjusted = colors.LinearSegmentedColormap.from_list('trunc(' + cmap.name +', ' + str(0) + ',' + str(1) + ')', cmap(np.linspace(0,1,100)))
scaled = np.reshape(selection[:,0:firstLabelIndex], (selection.shape[0], reduceActivationsBy, reduceActivationsTo))
scaled = scaled.mean(axis=1)
plt.imshow(scaled, cmap=cmap_adjusted, interpolation='none')
ax = plt.gca()
# ax.pcolormesh(scaled, cmap=plt.get_cmap('afmhot'))
ax.tick_params(axis='both', which='major', bottom=False, top=False, left=False, right=False)
ticks = np.arange(activationsPerLabel * 0.5,selection.shape[0],activationsPerLabel)
ax.set_yticks(ticks)
ax.set_yticklabels(tickLabels)
plt.savefig(plotFileName, bbox_inches='tight')
def generateKMeansSeries(activationsPath, labelIndexMappingPath, targetFolder):
if targetFolder.endswith('/') == False:
targetFolder += '/'
activations = utility.arrayFromFile(activationsPath)
labels = utility.getLabelStrings(labelIndexMappingPath)
neurons = activations.shape[1] - len(labels)
# test per label k-means
k = PER_LABEL_START
while k <= PER_LABEL_END:
runCounter = 0
while runCounter < RUNS_PER_TYPE:
logger.info("Calculating per label KMeans with k = " + str(k) +
".")
currentTarget = targetFolder + "perLabel_" + str(
k) + "/run_" + str(runCounter) + "/"
if not os.path.exists(os.path.dirname(currentTarget)):
os.makedirs(os.path.dirname(currentTarget))
c, i = kMeans_per_label.findKMeansPerLabel(activations, k,
len(labels),
currentTarget,
labelIndexMappingPath)
plot_cluster.plotClusters(kMeans_core.cleanUp(c), i, neurons,
currentTarget, labels[:len(labels)])
runCounter += 1
k += 1
# test mixed k-means
k = MIXED_START
while k <= MIXED_END:
runCounter = 0
while runCounter < RUNS_PER_TYPE:
logger.info("Calculating mixed KMeans with k = " + str(k) + ".")
currentTarget = targetFolder + "mixed_" + str(k) + "/run_" + str(
runCounter) + "/"
if not os.path.exists(os.path.dirname(currentTarget)):
os.makedirs(os.path.dirname(currentTarget))
[c, i] = kMeans_mixed.findKMeans(activations, k, len(labels),
currentTarget)
plot_cluster.plotClusters(kMeans_core.cleanUp(c), i,
activations.shape[1] - len(labels),
currentTarget, labels[:len(labels)])
runCounter += 1
k += MIXED_STEP
def generateKMeansSeries(activationsPath, labelIndexMappingPath, targetFolder):
if targetFolder.endswith("/") == False:
targetFolder += "/"
activations = utility.arrayFromFile(activationsPath)
labels = utility.getLabelStrings(labelIndexMappingPath)
neurons = activations.shape[1] - len(labels)
# test per label k-means
k = PER_LABEL_START
while k <= PER_LABEL_END:
runCounter = 0
while runCounter < RUNS_PER_TYPE:
logger.info("Calculating per label KMeans with k = " + str(k) + ".")
currentTarget = targetFolder + "perLabel_" + str(k) + "/run_" + str(runCounter) + "/"
if not os.path.exists(os.path.dirname(currentTarget)):
os.makedirs(os.path.dirname(currentTarget))
c, i = kMeans_per_label.findKMeansPerLabel(
activations, k, len(labels), currentTarget, labelIndexMappingPath
)
plot_cluster.plotClusters(kMeans_core.cleanUp(c), i, neurons, currentTarget, labels[: len(labels)])
runCounter += 1
k += 1
# test mixed k-means
k = MIXED_START
while k <= MIXED_END:
runCounter = 0
while runCounter < RUNS_PER_TYPE:
logger.info("Calculating mixed KMeans with k = " + str(k) + ".")
currentTarget = targetFolder + "mixed_" + str(k) + "/run_" + str(runCounter) + "/"
if not os.path.exists(os.path.dirname(currentTarget)):
os.makedirs(os.path.dirname(currentTarget))
[c, i] = kMeans_mixed.findKMeans(activations, k, len(labels), currentTarget)
plot_cluster.plotClusters(
kMeans_core.cleanUp(c), i, activations.shape[1] - len(labels), currentTarget, labels[: len(labels)]
)
runCounter += 1
k += MIXED_STEP
utility.plotConfusionMatrix(confusionMatrix, labels, currentTargetPath + "confusion.pdf")
kCounter += 1
results = np.array(results)
logger.info('Writing file ' + targetPath + "overview.csv")
np.savetxt( targetPath + "overview.csv", results, delimiter=',')
utility.plotKMeansOverview(results, targetPath + "overview.pdf", False)
if __name__ == '__main__':
if len(sys.argv) != 5 and len(sys.argv) != 6:
logger.info("Please provide as argument:")
logger.info("1) Path to training activations (*.npy).")
logger.info("2) Path to test activations (*.npy).")
logger.info("3) Path to label mapping.")
logger.info("4) Path for target folder.")
logger.info("5) Neighbour relations (optional)")
sys.exit();
trainingActivations = utility.arrayFromFile(sys.argv[1])
testActivations = utility.arrayFromFile(sys.argv[2])
labels = utility.getLabelStrings(sys.argv[3])
nn = None
if len(sys.argv) == 6:
nn = utility.arrayFromFile(sys.argv[5])
runTests(trainingActivations, testActivations, labels, sys.argv[4], nn)
logger.info("4) Target path for evaluation results.")
sys.exit();
labelIndexInfoPath = sys.argv[1]
labelIndexMappingPath = sys.argv[2]
logFilePath = sys.argv[3]
evaluationTargetPath = sys.argv[4]
if evaluationTargetPath.endswith('/') == False:
evaluationTargetPath += '/'
if not os.path.exists(os.path.dirname(evaluationTargetPath)) and os.path.dirname(evaluationTargetPath) != '':
os.makedirs(os.path.dirname(evaluationTargetPath))
plotTrainingLossAndAccuracy(logFilePath, evaluationTargetPath)
# sys.exit()
labels = utility.getLabelStrings(labelIndexMappingPath)
[confusionMatrix, meanAveragePrecision, overallCorrect, overallWrong] = testNeuralNet(labels, labelIndexInfoPath, evaluationTargetPath)
overviewPath = evaluationTargetPath + "overview.csv"
logger.info('Writing file ' + overviewPath)
np.savetxt(overviewPath, np.array([0,meanAveragePrecision, overallCorrect, overallWrong]), delimiter=',')
confusionMatrixPath = evaluationTargetPath + 'confusionMatrix.npy'
logger.info('Writing file ' + confusionMatrixPath)
with open(confusionMatrixPath, "w") as outputFile:
np.save(outputFile, confusionMatrix)
utility.plotConfusionMatrix(confusionMatrix, labels, evaluationTargetPath + 'confusionMatrix.pdf')
currentTargetPath + "confusion.pdf")
kCounter += 1
results = np.array(results)
logger.info('Writing file ' + targetPath + "overview.csv")
np.savetxt(targetPath + "overview.csv", results, delimiter=',')
utility.plotKMeansOverview(results, targetPath + "overview.pdf", False)
if __name__ == '__main__':
if len(sys.argv) != 5 and len(sys.argv) != 6:
logger.info("Please provide as argument:")
logger.info("1) Path to training activations (*.npy).")
logger.info("2) Path to test activations (*.npy).")
logger.info("3) Path to label mapping.")
logger.info("4) Path for target folder.")
logger.info("5) Neighbour relations (optional)")
sys.exit()
trainingActivations = utility.arrayFromFile(sys.argv[1])
testActivations = utility.arrayFromFile(sys.argv[2])
labels = utility.getLabelStrings(sys.argv[3])
nn = None
if len(sys.argv) == 6:
nn = utility.arrayFromFile(sys.argv[5])
runTests(trainingActivations, testActivations, labels, sys.argv[4], nn)
labelIndexInfoPath = sys.argv[1]
labelIndexMappingPath = sys.argv[2]
logFilePath = sys.argv[3]
evaluationTargetPath = sys.argv[4]
if evaluationTargetPath.endswith('/') == False:
evaluationTargetPath += '/'
if not os.path.exists(os.path.dirname(evaluationTargetPath)
) and os.path.dirname(evaluationTargetPath) != '':
os.makedirs(os.path.dirname(evaluationTargetPath))
plotTrainingLossAndAccuracy(logFilePath, evaluationTargetPath)
# sys.exit()
labels = utility.getLabelStrings(labelIndexMappingPath)
[confusionMatrix, meanAveragePrecision, overallCorrect,
overallWrong] = testNeuralNet(labels, labelIndexInfoPath,
evaluationTargetPath)
overviewPath = evaluationTargetPath + "overview.csv"
logger.info('Writing file ' + overviewPath)
np.savetxt(overviewPath,
np.array(
[0, meanAveragePrecision, overallCorrect, overallWrong]),
delimiter=',')
confusionMatrixPath = evaluationTargetPath + 'confusionMatrix.npy'
logger.info('Writing file ' + confusionMatrixPath)
with open(confusionMatrixPath, "w") as outputFile:
np.save(outputFile, confusionMatrix)
