def __init__(self,
feature_size,
type_hierarchy,
current_types,
level=0,
lambda_reg=0.1,
max_iter=5000,
threshold=0.1):
if level == 0:
self._svm = MulticlassSVM(feature_size, len(current_types),
lambda_reg, max_iter, 'ova')
else:
self._svm = MulticlassSVM(feature_size,
len(current_types) + 1, lambda_reg,
max_iter, 'ova')
self._typemapping = {} # map type_id to class_id in this level
self._classmapping = {} # map class_id to type_id in this level
self._children = {} # map type_id to subtype classifier if exits
self._level = level
self._threshold = threshold
class_id = 0
# add other class
if level != 0:
self._typemapping[-1] = class_id
self._classmapping[class_id] = -1
class_id += 1
for t in current_types:
self._typemapping[t] = class_id
self._classmapping[class_id] = t
# check if t has subtypes
if t in type_hierarchy:
self._children[t] = HierarchySVM(feature_size, type_hierarchy,
type_hierarchy[t], level + 1,
lambda_reg, max_iter)
class_id += 1
def run():
logging.getLogger().setLevel(logging.WARNING)
d = Dataset()
#d.use_images_in_folder("/home/simon/Datasets/ImageNet_Natural/images/")
#d.use_images_in_folder("/home/simon/Datasets/ICAO_german/")
d.use_images_in_folder("/home/simon/Datasets/desko_ids/images_unique/")
#d.use_images_in_folder("/home/simon/Datasets/croatianFishDataset-final/")
#d.use_images_in_folder("/home/jaeger/data/croatianFishDataset1-5Dir/")
d.create_labels_from_path()
d.fill_split_assignments(1)
#d.read_from_file("/home/simon/Datasets/CUB_200_2011/cropped_scaled_alex.txt","imagepaths","string")
#d.read_from_file("/home/simon/Datasets/CUB_200_2011/tr_ID.txt","split_assignments","int")
#d.read_from_file("/home/simon/Datasets/CUB_200_2011/labels.txt","labels","int")
c = Classification()
c.add_algorithm(Resize(512, 320))
# #c.add_algorithm(Noise('saltpepper',0.1))
p = ParallelAlgorithm()
#
p1 = AlgorithmPipeline()
p1.add_algorithm(HOG())
p1.add_algorithm(SpatialPyramid())
# #p1.add_algorithm(MinMaxNormalize())
p1.add_algorithm(NormNormalize())
p.add_pipeline(p1)
p2 = AlgorithmPipeline()
p2.add_algorithm(Resize(64, 32))
p2.add_algorithm(Colorname())
p2.add_algorithm(SpatialPyramid())
p2.add_algorithm(NormNormalize())
# #p2.add_algorithm(MinMaxNormalize())
p.add_pipeline(p2)
c.add_algorithm(p)
# #c.add_algorithm(MinMaxNormalize())
#c.add_algorithm(NormNormalize())
# c.add_algorithm(MeanCalculator())
#c.add_algorithm(Resize(32,24))
c.add_algorithm(MulticlassSVM())
# #c.train(d)
# #for path, gt_label in zip(d.imagepaths, d.labels):
# # logging.info("Predicted class for " + path + " is " + str(c.predict(path).data[0]) + " (GT: " + str(gt_label) + ")")
## Caffe features
#c.add_algorithm(Caffe("","","fc7"))
#c.add_algorithm(MulticlassSVM())
#with open('run_evaluation.py', 'r') as fin:
# print(fin.read())
mean_acc, mean_mAP = Evaluation.random_split_eval(
d, c, absolute_train_per_class=1, runs=1)
#mean_acc,mean_mAP = Evaluation.fixed_split_eval(d,c)
logging.warning("Total accuracy is " + str(mean_acc))
logging.warning("Total mAP is " + str(mean_mAP))
def __init__(self, feature_size, type_hierarchy, current_types, level=0, lambda_reg=0.1, max_iter=5000, threshold=0.1):
if level ==0 :
self._svm = MulticlassSVM(feature_size, len(current_types), lambda_reg, max_iter, 'ova')
else:
self._svm = MulticlassSVM(feature_size, len(current_types)+1, lambda_reg, max_iter, 'ova')
self._typemapping = {} # map type_id to class_id in this level
self._classmapping = {} # map class_id to type_id in this level
self._children = {} # map type_id to subtype classifier if exits
self._level = level
self._threshold = threshold
class_id = 0
# add other class
if level != 0:
self._typemapping[-1] = class_id
self._classmapping[class_id] = -1
class_id += 1
for t in current_types:
self._typemapping[t] = class_id
self._classmapping[class_id] = t
# check if t has subtypes
if t in type_hierarchy:
self._children[t] = HierarchySVM(feature_size, type_hierarchy, type_hierarchy[t], level+1, lambda_reg, max_iter)
class_id += 1
class HierarchySVM:
def __init__(self, feature_size, type_hierarchy, current_types, level=0, lambda_reg=0.1, max_iter=5000, threshold=0.1):
if level ==0 :
self._svm = MulticlassSVM(feature_size, len(current_types), lambda_reg, max_iter, 'ova')
else:
self._svm = MulticlassSVM(feature_size, len(current_types)+1, lambda_reg, max_iter, 'ova')
self._typemapping = {} # map type_id to class_id in this level
self._classmapping = {} # map class_id to type_id in this level
self._children = {} # map type_id to subtype classifier if exits
self._level = level
self._threshold = threshold
class_id = 0
# add other class
if level != 0:
self._typemapping[-1] = class_id
self._classmapping[class_id] = -1
class_id += 1
for t in current_types:
self._typemapping[t] = class_id
self._classmapping[class_id] = t
# check if t has subtypes
if t in type_hierarchy:
self._children[t] = HierarchySVM(feature_size, type_hierarchy, type_hierarchy[t], level+1, lambda_reg, max_iter)
class_id += 1
def fit_em(self, train_x, train_y):
"""
row = [0]*len(x)
data = [1]*len(x)
train_x = list of list
train_y = list of list
:param x:
:param y:
:return:
"""
new_train_x = []
new_train_y = []
for i in xrange(len(train_y)):
x = train_x[i]
y = train_y[i]
flag = True
for l in y:
if l in self._typemapping:
flag = False
new_train_x.append(x)
new_train_y.append(self._typemapping[l])
if flag:
new_train_x.append(x)
new_train_y.append(0)
if len(new_train_y)>0:
self._svm.fit(new_train_x, new_train_y)
# train children svm
for child in self._children:
new_train_x = []
new_train_y = []
for i in xrange(len(train_y)):
x = train_x[i]
y = train_y[i]
if child in y:
new_train_x.append(x)
new_train_y.append(y)
print "Train child svm for label %d, example:#%d" % (child, len(new_train_y))
self._children[child].fit_em(new_train_x, new_train_y)
def predict_em(self, x):
labels = set()
c,score = self._svm.predict(x)
if self._classmapping[c] == -1:
return labels
else:
label = self._classmapping[c]
if score>self._threshold:
labels.add(label)
if label in self._children:
sub_svm = self._children[label]
labels.update(sub_svm.predict_em(x))
elif self._level==0:
labels.add(label)
return labels
class HierarchySVM:
def __init__(self, feature_size, type_hierarchy, current_types, level=0, lambda_reg=0.1, max_iter=5000, threshold=0.1):
if level ==0 :
self._svm = MulticlassSVM(feature_size, len(current_types), lambda_reg, max_iter, 'ova')
else:
self._svm = MulticlassSVM(feature_size, len(current_types)+1, lambda_reg, max_iter, 'ova')
self._typemapping = {} # map type_id to class_id in this level
self._classmapping = {} # map class_id to type_id in this level
self._children = {} # map type_id to subtype classifier if exits
self._level = level
self._threshold = threshold
class_id = 0
# add other class
if level != 0:
self._typemapping[-1] = class_id
self._classmapping[class_id] = -1
class_id += 1
for t in current_types:
self._typemapping[t] = class_id
self._classmapping[class_id] = t
# check if t has subtypes
if t in type_hierarchy:
self._children[t] = HierarchySVM(feature_size, type_hierarchy, type_hierarchy[t], level+1, lambda_reg, max_iter)
class_id += 1
def fit(self, train_x, train_y):
"""
row = [0]*len(x)
data = [1]*len(x)
train_x = list of list
train_y = list of list
:param x:
:param y:
:return:
"""
new_train_x = []
new_train_y = []
for i in xrange(len(train_y)):
x = train_x[i]
y = train_y[i]
flag = True
for l in y:
if l in self._typemapping:
flag = False
new_train_x.append(x)
new_train_y.append(self._typemapping[l])
if flag:
new_train_x.append(x)
new_train_y.append(0)
if len(new_train_y)>0:
self._svm.fit(new_train_x, new_train_y)
# train children svm
for child in self._children:
new_train_x = []
new_train_y = []
for i in xrange(len(train_y)):
x = train_x[i]
y = train_y[i]
if child in y:
new_train_x.append(x)
new_train_y.append(y)
print "Train child svm for label %d, example:#%d" % (child, len(new_train_y))
self._children[child].fit(new_train_x, new_train_y)
def predict(self, x):
labels = set()
c,score = self._svm.predict(x)
if self._classmapping[c] == -1:
return labels
else:
label = self._classmapping[c]
if score>self._threshold:
labels.add(label)
if label in self._children:
sub_svm = self._children[label]
labels.update(sub_svm.predict(x))
elif self._level==0:
labels.add(label)
return labels
