def crossValidate(train_y, train_X):
""" Cross validate to get optimal parameters """
# scale data
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1, 1))
X_scaledArr = min_max_scaler.fit_transform(train_X)
X_scaled = X_scaledArr.tolist()
# write to svm format file
outputPath = '.\\cv'
fileName = 'train_data'
svmTool = SvmTool()
svmTool.write2SVMFormat(outputPath, fileName, X_scaled, train_y)
def crossValidate(train_y, train_X):
""" Cross validate to get optimal parameters """
# scale data
min_max_scaler = preprocessing.MinMaxScaler(feature_range=(-1, 1))
X_scaledArr = min_max_scaler.fit_transform(train_X)
X_scaled = X_scaledArr.tolist()
# write to svm format file
outputPath = '.\\cv'
fileName = 'train_data'
svmTool = SvmTool()
svmTool.write2SVMFormat(outputPath, fileName, X_scaled, train_y)
def main():
imageDatasetDir = config.imageDatasetDir
# Select part of the whole categories except 'BACKGROUND_Google'
categories = os.listdir(imageDatasetDir)
categories = categories[:config.numCategories]
allFeatures = np.array([])
## Step.1 Data loading and _features extraction
# Get _features from training data over all categories
print "Data loading and feature extraction ..."
trainImageData = []
testImageData = []
for category in categories:
categoryPath = os.path.join(imageDatasetDir, category)
allData = os.listdir(categoryPath)
numData = len(allData) # number of all data of the category
numTrainData = int(numData * config.percentageTrainData) # number of training data
trainData = allData[:numTrainData]
testData = allData[numTrainData:]
# Train data loading
for data in trainData:
filePath = os.path.join(categoryPath, data)
imageData = ImageData(filePath)
imageData.extractFeatures()
imageData.className = category
imageData.classId = categories.index(category)
trainImageData.append(imageData)
if allFeatures.size == 0:
allFeatures = imageData.features
else:
allFeatures = np.vstack((allFeatures, imageData.features))
# Test data loading
for data in testData:
filePath = os.path.join(categoryPath, data)
imageData = ImageData(filePath)
imageData.extractFeatures()
imageData.className = category
imageData.classId = categories.index(category)
testImageData.append(imageData)
## Step.2 Codebook generation
print "Codebook generation ..."
bovw = BagOfWords()
if os.path.exists('codebook.npy'):
codebook = np.load('codebook.npy')
bovw.codebook = codebook
else:
bovw.generateCodebook(allFeatures)
## Step.3 Feature encoding for train data
train_y = []
train_X = []
for imageData in trainImageData:
# Feature encoding, pooling, normalization
imageData.generateFinalFeatures(bovw)
# Format train data
train_y.append(imageData.classId)
train_X.append(imageData._finalFeatures)
# Cross validation
if config.is_cv:
# cross validation
crossValidate(train_y, train_X)
## Step.4 Classification
# Learning using SVM
svmTool = SvmTool()
print "Model learning ..."
svmTool.learnModel(train_y, train_X)
# Feature encoding for test data and classify data using learned model
print "Classifying ..."
numCorrect = 0
for imageData in testImageData:
# Feature encoding, pooling, normalization
imageData.generateFinalFeatures(bovw)
# Format train data
test_y = [imageData.classId]
test_X = [imageData.finalFeatures]
p_label, _ = svmTool.doPredication(test_y, test_X)
predClassId = int(p_label[0])
predClassName = categories[predClassId]
# Write test image to predicated category
predClassPath = os.path.join(config.predctDir, predClassName)
if not os.path.exists(predClassPath):
os.makedirs(predClassPath)
imageName = imageData.className + '_' + os.path.basename(imageData.filepath)
predFilePath = os.path.join(predClassPath, imageName)
cv2.imwrite(predFilePath, imageData.image)
if predClassId == imageData.classId:
numCorrect += 1
# Calculate results
accuracy = numCorrect / float(len(testImageData))
print "accuracy: ", accuracy
def main():
actionDatabaseDir = config.actionDatabaseDir
categories = config.actionCateogory
## Step.1 Data loading and features extraction
# Get features from training data over all categories
print "Data loading and feature extraction ..."
allFeatures = np.array([])
trainActionSequence = []
testActionSequence = []
if os.path.exists('allFeatures.npy') \
and os.path.exists('trainActionSequence.npy') \
and os.path.exists('testActionSequence.npy'):
allFeatures = np.load('allFeatures.npy')
trainActionSequence = np.load('trainActionSequence.npy')
testActionSequence = np.load('testActionSequence.npy')
else:
for category in categories:
categoryPath = os.path.join(actionDatabaseDir, category)
allData = os.listdir(categoryPath)
# Train data and test data loading
for data in allData:
filePath = os.path.join(categoryPath, data)
actionSequence = ActionSequence(filePath)
actionSequence.extractStip() # extract STIP
subject = actionSequence.subject
if subject in config.trainDataSubjects:
trainActionSequence.append(actionSequence)
if allFeatures.size == 0:
allFeatures = actionSequence.stipFeatures
else:
allFeatures = np.vstack((allFeatures,
actionSequence.stipFeatures))
elif subject in config.testDataSubjects:
testActionSequence.append(actionSequence)
np.save('allFeatures', allFeatures)
np.save('trainActionSequence', trainActionSequence)
np.save('testActionSequence', testActionSequence)
## Step.2 Codebook generation
print "Codebook generation ..."
bovw = BagOfWords(featureEncodingMethod = 'sparse-coding',
poolingMethod = 'max-pooling',
normalizationMethod = 'L2-norm')
# bovw = BagOfWords(featureEncodingMethod = 'vector-quantization',
# poolingMethod = 'sum-pooling',
# normalizationMethod = 'L1-norm')
if os.path.exists('codebook.npy'):
codebook = np.load('codebook.npy')
bovw.codebook = codebook
else:
bovw.generateCodebook(allFeatures)
np.save('codebook', bovw.codebook)
## Step.3 Feature encoding for train data
train_y = []
train_X = []
for actionSequence in trainActionSequence:
# Feature encoding, pooling, normalization
actionSequence.generateFinalFeatures(bovw)
# Format train data
train_y.append(actionSequence.categoryId)
train_X.append(actionSequence.finalFeatures)
# Cross validation
if config.is_cv:
# cross validation
crossValidate(train_y, train_X)
## Step.4 Classification
# Learning using SVM
svmTool = SvmTool()
print "Model learning ..."
svmTool.learnModel(train_y, train_X)
# Feature encoding for test data and classify data using learned model
print "Predicating ..."
numCorrect = 0
trueLabels = []
testPredLabels = []
removeall(config.predctDir)
for actionSequence in testActionSequence:
# Feature encoding, pooling, normalization
actionSequence.generateFinalFeatures(bovw)
# Format train data
test_y = [actionSequence.categoryId]
test_X = [actionSequence.finalFeatures]
p_label, _ = svmTool.doPredication(test_y, test_X)
predCagtegoryId = int(p_label[0])
predCategoryName = categories[predCagtegoryId]
# Write predicated action to predicated category
predCategoryPath = os.path.join(config.predctDir, predCategoryName)
if not os.path.exists(predCategoryPath):
os.makedirs(predCategoryPath)
predFilePath = os.path.join(predCategoryPath, actionSequence.filename)
f = open(predFilePath, 'w')
f.close()
if predCagtegoryId == actionSequence.categoryId:
numCorrect += 1
trueLabels.append(actionSequence.categoryName)
testPredLabels.append(predCategoryName)
# Calculate results
accuracy = numCorrect / float(len(testActionSequence))
print "accuracy: ", accuracy
# Plot confusion matrix
saveFilename = 'confusion_matrix.png'
plotConfusionMatrix(trueLabels, testPredLabels,
saveFilename, normalization = False)
saveFilename = 'confusion_matrix_norm.png'
plotConfusionMatrix(trueLabels, testPredLabels,
saveFilename, normalization = True)