def torch_predict(V, W, classId, XlT, XuT, patClassIdTest, gama = 1, oper = 'min'):
"""
GFMM classifier (test routine). Implemented by Pytorch
result = predict(V,W,classId,XlT,XuT,patClassIdTest,gama,oper)
INPUT
V Tested model hyperbox lower bounds
W Tested model hyperbox upper bounds
classId Input data (hyperbox) class labels (crisp)
XlT Test data lower bounds (rows = objects, columns = features)
XuT Test data upper bounds (rows = objects, columns = features)
patClassIdTest Test data class labels (crisp)
gama Membership function slope (default: 1)
oper Membership calculation operation: 'min' or 'prod' (default: 'min')
OUTPUT
result A object with Bunch datatype containing all results as follows:
+ summis Number of misclassified objects
+ misclass Binary error map
+ sumamb Number of objects with maximum membership in more than one class
+ out Soft class memberships
+ mem Hyperbox memberships
"""
if len(XlT.size()) == 1:
XlT = XlT.reshape(1, -1)
if len(XuT.size()) == 1:
XuT = XuT.reshape(1, -1)
#initialization
yX = XlT.size(0)
isUsingGPU = False
if is_Have_GPU and (W.size(0) * W.size(1) >= GPU_Computing_Threshold or XlT.size(1) >= GPU_Computing_Threshold):
V = V.cuda()
W = W.cuda()
classId = classId.cuda()
XlT = XlT.cuda()
XuT = XuT.cuda()
patClassIdTest = patClassIdTest.cuda()
misclass = torch.cuda.FloatTensor(yX).fill_(0)
isUsingGPU = True
els = torch.arange(yX).cuda()
else:
misclass = torch.zeros(yX)
els = torch.arange(yX)
# classifications
for i in els:
if isUsingGPU == True:
mem = gpu_memberG(XlT[i, :], XuT[i, :], V, W, gama, oper)
else:
mem = torch_memberG(XlT[i, :], XuT[i, :], V, W, gama, oper) # calculate memberships for all hyperboxes
bmax = mem.max() # get max membership value
maxVind = mem == bmax # get indexes of all hyperboxes with max membership
if bmax == 0:
print('zero maximum membership value') # this is probably bad...
misclass[i] = 1
else:
if len(torch.unique(classId[maxVind])) > 1:
misclass[i] = 1
else:
if (torch.any(classId[maxVind] == patClassIdTest[i]) == 1) or (patClassIdTest[i] == UNLABELED_CLASS):
misclass[i] = 0
else:
misclass[i] = 1
# results
summis = torch.sum(misclass)
result = Bunch(summis = summis, misclass = misclass)
return result
def torch_predict(V, W, classId, XlT, XuT, patClassIdTest, gama=1, oper='min'):
"""
GFMM classifier (test routine). Implemented by Pytorch
result = predict(V,W,classId,XlT,XuT,patClassIdTest,gama,oper)
INPUT
V Tested model hyperbox lower bounds
W Tested model hyperbox upper bounds
classId Input data (hyperbox) class labels (crisp)
XlT Test data lower bounds (rows = objects, columns = features)
XuT Test data upper bounds (rows = objects, columns = features)
patClassIdTest Test data class labels (crisp)
gama Membership function slope (default: 1)
oper Membership calculation operation: 'min' or 'prod' (default: 'min')
OUTPUT
result A object with Bunch datatype containing all results as follows:
+ summis Number of misclassified objects
+ misclass Binary error map
+ sumamb Number of objects with maximum membership in more than one class
+ out Soft class memberships
+ mem Hyperbox memberships
"""
#initialization
yX = XlT.size(0)
isUsingGPU = False
if is_Have_GPU and (W.size(0) * W.size(1) >= GPU_Computing_Threshold
or XlT.size(1) >= GPU_Computing_Threshold):
V = V.cuda()
W = W.cuda()
classId = classId.cuda()
XlT = XlT.cuda()
XuT = XuT.cuda()
patClassIdTest = patClassIdTest.cuda()
misclass = torch.cuda.FloatTensor(yX).fill_(0)
classes = torch.unique(classId)
noClasses = classes.size(0)
ambiguity = torch.cuda.FloatTensor(yX, 1).fill_(0)
mem = torch.cuda.FloatTensor(yX, V.size(0)).fill_(0)
out = torch.cuda.FloatTensor(yX, noClasses).fill_(0)
isUsingGPU = True
els = torch.arange(yX).cuda()
else:
classes = torch.unique(classId)
misclass = torch.zeros(yX)
noClasses = classes.size(0)
ambiguity = torch.zeros((yX, 1))
mem = torch.zeros((yX, V.size(0)))
out = torch.zeros((yX, noClasses))
els = torch.arange(yX)
# classifications
for i in els:
if isUsingGPU == True:
mem[i, :] = gpu_memberG(XlT[i, :], XuT[i, :], V, W, gama, oper)
else:
mem[i, :] = torch_memberG(
XlT[i, :], XuT[i, :], V, W, gama,
oper) # calculate memberships for all hyperboxes
bmax = mem[i, :].max() # get max membership value
maxVind = torch.nonzero(mem[
i, :] == bmax) # get indexes of all hyperboxes with max membership
for j in torch.arange(noClasses):
out[i, j] = mem[i, classId == classes[j]].max(
) # get max memberships for each class
ambiguity[i, :] = torch.sum(out[
i, :] == bmax) # number of different classes with max membership
if bmax == 0:
print('zero maximum membership value') # this is probably bad...
misclass[i] = ~(torch.any(classId[maxVind] == patClassIdTest[i]) |
(patClassIdTest[i] == 0))
# results
sumamb = torch.sum(ambiguity[:, 0] > 1)
summis = torch.sum(misclass)
result = Bunch(summis=summis,
misclass=misclass,
sumamb=sumamb,
out=out,
mem=mem)
return result
def fit(self, X_l, X_u, patClassId):
"""
X_l Input data lower bounds (rows = objects, columns = features)
X_u Input data upper bounds (rows = objects, columns = features)
patClassId Input data class labels (crisp)
"""
if self.isNorm == True:
X_l, X_u = self.dataPreprocessing(X_l, X_u)
if isinstance(X_l, torch.Tensor) == False:
X_l = torch.from_numpy(X_l).float()
X_u = torch.from_numpy(X_u).float()
patClassId = torch.from_numpy(patClassId).long()
time_start = time.perf_counter()
isUsingGPU = False
if is_Have_GPU and X_l.size(0) * X_l.size(1) >= GPU_Computing_Threshold:
self.V = X_l.cuda()
self.W = X_u.cuda()
self.classId = patClassId.cuda()
isUsingGPU = True
else:
self.V = X_l
self.W = X_u
self.classId = patClassId
# yX, xX = X_l.size()
# if len(self.cardin) == 0 or len(self.clusters) == 0:
# self.cardin = np.ones(yX)
# self.clusters = np.empty(yX, dtype=object)
# for i in range(yX):
# self.clusters[i] = np.array([i], dtype = np.int32)
#
if self.isDraw:
mark_col = np.array(['r', 'g', 'b', 'y', 'c', 'm', 'k'])
drawing_canvas = self.initializeCanvasGraph("GFMM - AGGLO-SM-Fast version")
# plot initial hyperbox
Vt, Wt = self.pcatransform()
color_ = np.empty(len(self.classId), dtype = object)
for c in range(len(self.classId)):
color_[c] = mark_col[self.classId[c]]
drawbox(Vt, Wt, drawing_canvas, color_)
self.delay()
# training
isTraining = True
while isTraining:
isTraining = False
# calculate class masks
yX, xX = self.V.size()
labList = torch.unique(self.classId)
if isUsingGPU == False:
clMask = torch.zeros((yX, len(labList)), dtype=torch.uint8)
else:
clMask = torch.cuda.ByteTensor(yX, len(labList)).fill_(0)
for i in range(len(labList)):
clMask[:, i] = self.classId == labList[i]
# calculate pairwise memberships *ONLY* within each class (faster!)
if isUsingGPU == False:
b = torch.zeros((yX, yX))
else:
b = torch.cuda.FloatTensor(yX, yX).fill_(0)
if isUsingGPU:
els = torch.arange(len(labList)).cuda()
else:
els = torch.arange(len(labList))
for i in els:
Vi = self.V[clMask[:, i]] # get bounds of patterns with class label i
Wi = self.W[clMask[:, i]]
clSize = torch.sum(clMask[:, i]) # get number of patterns of class i
clIdxs = torch.nonzero(clMask[:, i])[:, 0] # get position of patterns with class label i in the training set
if self.simil == 'short':
for j in range(clSize):
if isUsingGPU == False:
b[clIdxs[j], clIdxs] = torch_memberG(Wi[j], Vi[j], Vi, Wi, self.gamma, self.oper)
else:
b[clIdxs[j], clIdxs] = gpu_memberG(Wi[j], Vi[j], Vi, Wi, self.gamma, self.oper)
elif self.simil == 'long':
for j in range(clSize):
if isUsingGPU == False:
b[clIdxs[j], clIdxs] = torch_memberG(Vi[j], Wi[j], Wi, Vi, self.gamma, self.oper)
else:
b[clIdxs[j], clIdxs] = gpu_memberG(Vi[j], Wi[j], Wi, Vi, self.gamma, self.oper)
else:
for j in range(clSize):
if isUsingGPU == False:
b[clIdxs[j], clIdxs] = torch_memberG(Vi[j], Wi[j], Vi, Wi, self.gamma, self.oper)
else:
b[clIdxs[j], clIdxs] = gpu_memberG(Vi[j], Wi[j], Vi, Wi, self.gamma, self.oper)
if yX == 1:
maxb = torch.FloatTensor([])
else:
maxb = self.torch_splitSimilarityMaxtrix(b, self.sing, False, isUsingGPU)
if len(maxb) > 0:
maxb = maxb[(maxb[:, 2] >= self.bthres), :]
if len(maxb) > 0:
# sort maxb in the decending order following the last column
values, idx_smaxb = torch.sort(maxb[:, 2], descending=True)
maxb = torch.cat((maxb[idx_smaxb, 0].reshape(-1, 1), maxb[idx_smaxb, 1].reshape(-1, 1), maxb[idx_smaxb, 2].reshape(-1, 1)), dim=1)
#maxb = maxb[idx_smaxb]
while len(maxb) > 0:
curmaxb = maxb[0, :] # current position handling
# calculate new coordinates of curmaxb(0)-th hyperbox by including curmaxb(1)-th box, scrap the latter and leave the rest intact
newV = torch.cat((self.V[0:curmaxb[0].long(), :], torch.min(self.V[curmaxb[0].long(), :], self.V[curmaxb[1].long(), :]).reshape(1, -1), self.V[curmaxb[0].long() + 1:curmaxb[1].long(), :], self.V[curmaxb[1].long() + 1:, :]), dim=0)
newW = torch.cat((self.W[0:curmaxb[0].long(), :], torch.max(self.W[curmaxb[0].long(), :], self.W[curmaxb[1].long(), :]).reshape(1, -1), self.W[curmaxb[0].long() + 1:curmaxb[1].long(), :], self.W[curmaxb[1].long() + 1:, :]), dim=0)
newClassId = torch.cat((self.classId[0:curmaxb[1].long()], self.classId[curmaxb[1].long() + 1:]))
#print('Type newV = ', newV.type())
# adjust the hyperbox if no overlap and maximum hyperbox size is not violated
if (not torch_isOverlap(newV, newW, curmaxb[0].long(), newClassId, isUsingGPU)) and (((newW[curmaxb[0].long()] - newV[curmaxb[0].long()]) <= self.teta).all() == True):
isTraining = True
self.V = newV
self.W = newW
self.classId = newClassId
# self.cardin[int(curmaxb[0])] = self.cardin[int(curmaxb[0])] + self.cardin[int(curmaxb[1])]
# self.cardin = np.append(self.cardin[0:int(curmaxb[1])], self.cardin[int(curmaxb[1]) + 1:])
#
# self.clusters[int(curmaxb[0])] = np.append(self.clusters[int(curmaxb[0])], self.clusters[int(curmaxb[1])])
# self.clusters = np.append(self.clusters[0:int(curmaxb[1])], self.clusters[int(curmaxb[1]) + 1:])
#
# remove joined pair from the list as well as any pair with lower membership and consisting of any of joined boxes
mask = (maxb[:, 0] != curmaxb[0]) & (maxb[:, 1] != curmaxb[0]) & (maxb[:, 0] != curmaxb[1]) & (maxb[:, 1] != curmaxb[1]) & (maxb[:, 2] >= curmaxb[2])
maxb = maxb[mask, :]
# update indexes to accomodate removed hyperbox
# indices of V and W larger than curmaxb(1,2) are decreased 1 by the element whithin the location curmaxb(1,2) was removed
if len(maxb) > 0:
maxb[maxb[:, 0] > curmaxb[1], 0] = maxb[maxb[:, 0] > curmaxb[1], 0] - 1
maxb[maxb[:, 1] > curmaxb[1], 1] = maxb[maxb[:, 1] > curmaxb[1], 1] - 1
if self.isDraw:
Vt, Wt = self.pcatransform()
color_ = np.empty(len(self.classId), dtype = object)
for c in range(len(self.classId)):
color_[c] = mark_col[self.classId[c]]
drawing_canvas.cla()
drawbox(Vt, Wt, drawing_canvas, color_)
self.delay()
else:
maxb = maxb[1:, :] # scrap examined pair from the list
if isTraining == True and isUsingGPU == True and self.V.size(0) * self.V.size(1) < GPU_Computing_Threshold:
isUsingGPU = False
self.V = self.V.cpu()
self.W = self.W.cpu()
self.classId = self.classId.cpu()
time_end = time.perf_counter()
self.elapsed_training_time = time_end - time_start
return self
def fit(self, X_l, X_u, patClassId):
"""
Training the classifier
Xl Input data lower bounds (rows = objects, columns = features)
Xu Input data upper bounds (rows = objects, columns = features)
patClassId Input data class labels (crisp). patClassId[i] = UNLABELED_CLASS corresponds to an unlabeled item
"""
print('--Online Learning--')
if self.isNorm == True:
X_l, X_u = self.dataPreprocessing(X_l, X_u)
if isinstance(X_l, torch.Tensor) == False:
X_l = torch.from_numpy(X_l).float()
X_u = torch.from_numpy(X_u).float()
patClassId = torch.from_numpy(patClassId).long()
time_start = time.perf_counter()
yX, xX = X_l.size()
teta = self.teta
self.misclass = 1
isUsingGPU = False
while self.misclass > 0 and teta >= self.tMin:
# for each input sample
for i in range(yX):
if len(
self.V
) > 0 and is_Have_GPU and isUsingGPU == False and self.V.size(
0) * self.V.size(1) >= GPU_Computing_Threshold:
self.V = self.V.cuda()
self.W = self.W.cuda()
self.classId = self.classId.cuda()
isUsingGPU = True
if self.V.size(
0) == 0: # no model provided - starting from scratch
self.V = X_l[0].reshape(
1, -1) # torch.DoubleTensor(X_l[0]).to(device)
self.W = X_u[0].reshape(
1, -1) # torch.DoubleTensor(X_u[0]).to(device)
self.classId = torch.LongTensor([
patClassId[0]
]) # torch.DoubleTensor([patClassId[0]]).to(device)
else:
if isUsingGPU == False:
classOfX = patClassId[i]
else:
classOfX = patClassId[i].cuda()
id_lb_sameX = (self.classId == classOfX) | (
self.classId == UNLABELED_CLASS)
if len(torch.nonzero(id_lb_sameX)) > 0:
V_sameX = self.V[id_lb_sameX]
W_sameX = self.W[id_lb_sameX]
lb_sameX = self.classId[id_lb_sameX]
id_range = torch.arange(len(self.classId))
id_processing = id_range[id_lb_sameX]
if isUsingGPU == False:
Xl_cur = X_l[i]
Xu_cur = X_u[i]
b = torch_memberG(Xl_cur, Xu_cur, V_sameX, W_sameX,
self.gamma)
else:
Xl_cur = X_l[i].cuda()
Xu_cur = X_u[i].cuda()
b = gpu_memberG(Xl_cur, Xu_cur, V_sameX, W_sameX,
self.gamma)
bSort, index = torch.sort(b, descending=True)
if bSort[0] != 1 or (classOfX != lb_sameX[index[0]]
and classOfX != UNLABELED_CLASS):
adjust = False
for j in id_processing[index]:
# test violation of max hyperbox size and class labels
if (classOfX == self.classId[j]
or self.classId[j] == UNLABELED_CLASS
or classOfX == UNLABELED_CLASS
) and (
(torch.max(self.W[j], Xu_cur).float() -
torch.min(self.V[j], Xl_cur).float())
<= teta).all() == True:
# adjust the j-th hyperbox
self.V[j] = torch.min(self.V[j], Xl_cur)
self.W[j] = torch.max(self.W[j], Xu_cur)
indOfWinner = j
adjust = True
if classOfX != UNLABELED_CLASS and self.classId[
j] == UNLABELED_CLASS:
self.classId[j] = classOfX
break
# if i-th sample did not fit into any existing box, create a new one
if not adjust:
self.V = torch.cat(
(self.V, Xl_cur.reshape(1, -1)), 0)
self.W = torch.cat(
(self.W, Xu_cur.reshape(1, -1)), 0)
if isUsingGPU == False:
self.classId = torch.cat(
(self.classId,
torch.LongTensor([classOfX])), 0)
else:
self.classId = torch.cat(
(self.classId,
torch.cuda.LongTensor([classOfX])), 0)
elif self.V.size(0) > 1:
for ii in range(self.V.size(0)):
if ii != indOfWinner and self.classId[
ii] != self.classId[indOfWinner]:
caseDim = torch_hyperboxOverlapTest(
self.V, self.W, indOfWinner,
ii) # overlap test
if len(caseDim) > 0:
self.V, self.W = torch_hyperboxContraction(
self.V, self.W, caseDim, ii,
indOfWinner)
else:
# create new sample
if isUsingGPU == False:
self.V = torch.cat((self.V, X_l[i].reshape(1, -1)),
0)
self.W = torch.cat((self.W, X_u[i].reshape(1, -1)),
0)
self.classId = torch.cat(
(self.classId, torch.LongTensor(
[patClassId[i]])), 0)
else:
self.V = torch.cat(
(self.V, X_l[i].cuda().reshape(1, -1)), 0)
self.W = torch.cat(
(self.W, X_l[i].cuda().reshape(1, -1)), 0)
self.classId = torch.cat(
(self.classId,
torch.cuda.LongTensor([patClassId[i]])), 0)
teta = teta * 0.9
if teta >= self.tMin:
result = torch_predict(self.V, self.W, self.classId, X_l, X_u,
patClassId, self.gamma, self.oper)
self.misclass = result.summis
time_end = time.perf_counter()
self.elapsed_training_time = time_end - time_start
return self
def fit(self, X_l, X_u, patClassId):
"""
Xl Input data lower bounds (rows = objects, columns = features)
Xu Input data upper bounds (rows = objects, columns = features)
patClassId Input data class labels (crisp)
"""
if self.isNorm == True:
X_l, X_u = self.dataPreprocessing(X_l, X_u)
if isinstance(X_l, torch.Tensor) == False:
X_l = torch.from_numpy(X_l).float()
X_u = torch.from_numpy(X_u).float()
patClassId = torch.from_numpy(patClassId).long()
time_start = time.perf_counter()
isUsingGPU = False
if is_Have_GPU and X_l.size(0) * X_l.size(1) >= GPU_Computing_Threshold:
self.V = X_l.cuda()
self.W = X_u.cuda()
self.classId = patClassId.cuda()
isUsingGPU = True
else:
self.V = X_l
self.W = X_u
self.classId = patClassId
# yX, xX = X_l.shape
# if len(self.cardin) == 0 or len(self.clusters) == 0:
# self.cardin = np.ones(yX)
# self.clusters = np.empty(yX, dtype=object)
# for i in range(yX):
# self.clusters[i] = np.array([i], dtype = np.int64)
if self.isDraw:
mark_col = np.array(['r', 'g', 'b', 'y', 'c', 'm', 'k'])
drawing_canvas = self.initializeCanvasGraph("GFMM - AGGLO-2")
# plot initial hyperbox
Vt, Wt = self.pcatransform()
color_ = np.empty(len(self.classId), dtype=object)
for c in range(len(self.classId)):
color_[c] = mark_col[self.classId[c]]
boxes = drawbox(Vt, Wt, drawing_canvas, color_)
self.delay()
hyperboxes = list(boxes)
# training
isTraining = True
while isTraining:
isTraining = False
k = 0 # input pattern index
while k < len(self.classId):
if self.simil == 'short':
if isUsingGPU == False:
b = torch_memberG(self.W[k], self.V[k], self.V, self.W,
self.gamma, self.oper, isUsingGPU)
else:
b = gpu_memberG(self.W[k], self.V[k], self.V, self.W,
self.gamma, self.oper)
elif self.simil == 'long':
if isUsingGPU == False:
b = torch_memberG(self.V[k], self.W[k], self.W, self.V,
self.gamma, self.oper, isUsingGPU)
else:
b = gpu_memberG(self.V[k], self.W[k], self.W, self.V,
self.gamma, self.oper)
else:
b = torch_asym_similarity_one_many(self.V[k], self.W[k],
self.V, self.W,
self.gamma, self.sing,
self.oper, isUsingGPU)
sortB, indB = torch.sort(b, descending=True)
maxB = sortB[sortB >=
self.bthres] # apply membership threshold
if len(maxB) > 0:
indmaxB = indB[sortB >= self.bthres]
# remove self-membership
maxB = maxB[indmaxB != k]
indmaxB = indmaxB[indmaxB != k]
# remove memberships to boxes from other classes
# idx_other_classes = np.where(np.logical_and(self.classId[indmaxB] != self.classId[k], self.classId[indmaxB] != 0))
#idx_same_classes = np.where(np.logical_or(self.classId[indmaxB] == self.classId[k], self.classId[indmaxB] == 0))
#idx_same_classes = np.where(self.classId[indmaxB] == self.classId[k])[0] # np.logical_or(self.classId[indmaxB] == self.classId[k], self.classId[indmaxB] == 0)
#maxB = np.delete(maxB, idx_other_classes)
idx_same_classes = (self.classId[indmaxB]
== self.classId[k]) | (
self.classId[indmaxB] == 0)
maxB = maxB[idx_same_classes]
# leaving memeberships to unlabelled boxes
indmaxB = indmaxB[idx_same_classes]
# if len(maxB) > 30: # trim the set of memberships to speedup processing
# maxB = maxB[0:30]
# indmaxB = indmaxB[0:30]
if isUsingGPU == True:
kMat = torch.cuda.LongTensor([k]).expand(
indmaxB.size(0))
else:
kMat = torch.LongTensor([k]).expand(indmaxB.size(0))
pairewise_maxb = torch.cat(
(torch.min(kMat, indmaxB).reshape(
-1, 1).float(), torch.max(kMat, indmaxB).reshape(
-1, 1).float(), maxB.reshape(-1, 1)),
dim=1)
if isUsingGPU:
els = torch.arange(pairewise_maxb.size(0)).cuda()
else:
els = torch.arange(pairewise_maxb.size(0))
for i in els:
# calculate new coordinates of k-th hyperbox by including pairewise_maxb(i,1)-th box, scrap the latter and leave the rest intact
# agglomorate pairewise_maxb(i, 0) and pairewise_maxb(i, 1) by adjusting pairewise_maxb(i, 0)
# remove pairewise_maxb(i, 1) by getting newV from 1 -> pairewise_maxb(i, 0) - 1, new coordinates for pairewise_maxb(i, 0), from pairewise_maxb(i, 0) + 1 -> pairewise_maxb(i, 1) - 1, pairewise_maxb(i, 1) + 1 -> end
newV = torch.cat(
(self.V[:pairewise_maxb[i, 0].long()],
torch.min(
self.V[pairewise_maxb[i, 0].long()],
self.V[pairewise_maxb[i, 1].long()]).reshape(
1, -1),
self.V[pairewise_maxb[i, 0].long() +
1:pairewise_maxb[i, 1].long()],
self.V[pairewise_maxb[i, 1].long() + 1:]),
dim=0)
newW = torch.cat(
(self.W[:pairewise_maxb[i, 0].long()],
torch.max(
self.W[pairewise_maxb[i, 0].long()],
self.W[pairewise_maxb[i, 1].long()]).reshape(
1, -1),
self.W[pairewise_maxb[i, 0].long() +
1:pairewise_maxb[i, 1].long()],
self.W[pairewise_maxb[i, 1].long() + 1:]),
dim=0)
newClassId = torch.cat(
(self.classId[:pairewise_maxb[i, 1].long()],
self.classId[pairewise_maxb[i, 1].long() + 1:]))
# adjust the hyperbox if no overlap and maximum hyperbox size is not violated
# position of adjustment is pairewise_maxb[i, 0] in new bounds
if (not torch_isOverlap(
newV, newW, pairewise_maxb[i, 0].long(),
newClassId)) and (
((newW[pairewise_maxb[i, 0].long()] -
newV[pairewise_maxb[i, 0].long()]) <=
self.teta).all() == True):
self.V = newV
self.W = newW
self.classId = newClassId
# self.cardin[int(pairewise_maxb[i, 0])] = self.cardin[int(pairewise_maxb[i, 0])] + self.cardin[int(pairewise_maxb[i, 1])]
# #self.cardin = np.delete(self.cardin, int(pairewise_maxb[i, 1]))
# self.cardin = np.append(self.cardin[0:int(pairewise_maxb[i, 1])], self.cardin[int(pairewise_maxb[i, 1]) + 1:])
#
# self.clusters[int(pairewise_maxb[i, 0])] = np.append(self.clusters[int(pairewise_maxb[i, 0])], self.clusters[int(pairewise_maxb[i, 1])])
# #self.clusters = np.delete(self.clusters, int(pairewise_maxb[i, 1]))
# self.clusters = np.append(self.clusters[0:int(pairewise_maxb[i, 1])], self.clusters[int(pairewise_maxb[i, 1]) + 1:])
#
isTraining = True
if k != pairewise_maxb[
i,
0]: # position pairewise_maxb[i, 1] (also k) is removed, so next step should start from pairewise_maxb[i, 1]
k = k - 1
if self.isDraw:
try:
hyperboxes[pairewise_maxb[
i, 1].long()].remove()
hyperboxes[pairewise_maxb[
i, 0].long()].remove()
except:
print("No remove old hyperbox")
Vt, Wt = self.pcatransform()
box_color = 'k'
if self.classId[pairewise_maxb[
i, 0].long()] < len(mark_col):
box_color = mark_col[self.classId[
pairewise_maxb[i, 0].long()]]
box = drawbox(
np.asmatrix(Vt[pairewise_maxb[i,
0].long()]),
np.asmatrix(Wt[pairewise_maxb[i,
0].long()]),
drawing_canvas, box_color)
self.delay()
hyperboxes[pairewise_maxb[i,
0].long()] = box[0]
hyperboxes.remove(
hyperboxes[pairewise_maxb[i, 1].long()])
break # if hyperbox adjusted there's no need to look at other hyperboxes
k = k + 1
if isTraining == True and isUsingGPU == True and self.V.size(
0) * self.V.size(1) < GPU_Computing_Threshold:
isUsingGPU = False
self.V = self.V.cpu()
self.W = self.W.cpu()
self.classId = self.classId.cpu()
time_end = time.perf_counter()
self.elapsed_training_time = time_end - time_start
return self
def fit(self, X_l, X_u, patClassId):
"""
Training the classifier
Xl Input data lower bounds (rows = objects, columns = features)
Xu Input data upper bounds (rows = objects, columns = features)
patClassId Input data class labels (crisp). patClassId[i] = 0 corresponds to an unlabeled item
"""
print('--Online Learning--')
if self.isNorm == True:
X_l, X_u = self.dataPreprocessing(X_l, X_u)
if isinstance(X_l, torch.Tensor) == False:
# X_l = torch.cuda.FloatTensor(X_l)
# X_u = torch.cuda.FloatTensor(X_u)
# patClassId = torch.cuda.LongTensor(patClassId)
# print('Conver data')
# t1 = time.clock()
X_l = torch.from_numpy(X_l).float()
X_u = torch.from_numpy(X_u).float()
patClassId = torch.from_numpy(patClassId).long()
# t2 = time.clock()
# print('Finish Conver data: ', t2 - t1)
time_start = time.perf_counter()
yX, xX = X_l.size()
teta = self.teta
mark = np.array([
'*', 'o', 'x', '+', '.', ',', 'v', '^', '<', '>', '1', '2', '3',
'4', '8', 's', 'p', 'P', 'h', 'H', 'X', 'D', '|', '_'
])
mark_col = np.array(['r', 'g', 'b', 'y', 'c', 'm', 'k'])
listLines = list()
listInputSamplePoints = list()
if self.isDraw:
drawing_canvas = self.initializeCanvasGraph(
"GFMM - Online learning", xX)
if self.V.size(0) > 0:
# draw existed hyperboxes
color_ = np.array(['k'] * len(self.classId), dtype=object)
for c in range(len(self.classId)):
if self.classId[c] < len(mark_col):
color_[c] = mark_col[self.classId[c]]
hyperboxes = drawbox(self.V[:, 0:np.minimum(xX, 3)].numpy(),
self.W[:, 0:np.minimum(xX, 3)].numpy(),
drawing_canvas, color_)
listLines.extend(hyperboxes)
self.delay()
self.misclass = 1
isUsingGPU = False
while self.misclass > 0 and teta >= self.tMin:
# for each input sample
for i in range(yX):
if len(
self.V
) > 0 and is_Have_GPU and isUsingGPU == False and self.V.size(
0) * self.V.size(1) >= GPU_Computing_Threshold:
self.V = self.V.cuda()
self.W = self.W.cuda()
self.classId = self.classId.cuda()
isUsingGPU = True
# print('Sample: ', i)
# draw input samples
if self.isDraw:
if i == 0 and len(listInputSamplePoints) > 0:
# reset input point drawing
for point in listInputSamplePoints:
point.remove()
listInputSamplePoints.clear()
color_ = 'k'
if patClassId[i] < len(mark_col):
color_ = mark_col[patClassId[i]]
if (X_l[i, :] == X_u[i, :]).all():
marker_ = 'd'
if patClassId[i] < len(mark):
marker_ = mark[patClassId[i]]
if xX == 2:
inputPoint = drawing_canvas.plot(X_l[i, 0],
X_l[i, 1],
color=color_,
marker=marker_)
else:
inputPoint = drawing_canvas.plot([X_l[i, 0]],
[X_l[i, 1]],
[X_l[i, 2]],
color=color_,
marker=marker_)
#listInputSamplePoints.append(inputPoint)
else:
inputPoint = drawbox(
np.asmatrix(X_l[i, 0:np.minimum(xX, 3)].numpy()),
np.asmatrix(X_u[i, 0:np.minimum(xX, 3)].numpy()),
drawing_canvas, color_)
listInputSamplePoints.append(inputPoint[0])
self.delay()
if self.V.size(
0) == 0: # no model provided - starting from scratch
#print('Initial data')
#t1 = time.clock()
self.V = X_l[0].reshape(
1, -1) # torch.DoubleTensor(X_l[0]).to(device)
self.W = X_u[0].reshape(
1, -1) # torch.DoubleTensor(X_u[0]).to(device)
self.classId = torch.LongTensor([
patClassId[0]
]) # torch.DoubleTensor([patClassId[0]]).to(device)
#t2 = time.clock()
#print('Finish Initial data: ', t2 - t1)
if self.isDraw == True:
# draw hyperbox
box_color = 'k'
if patClassId[0] < len(mark_col):
box_color = mark_col[patClassId[0]]
hyperbox = drawbox(
np.asmatrix(self.V[0,
0:np.minimum(xX, 3)].numpy()),
np.asmatrix(self.W[0,
0:np.minimum(xX, 3)].numpy()),
drawing_canvas, box_color)
listLines.append(hyperbox[0])
self.delay()
else:
# print('V === ', self.V)
# print('W === ', self.W)
# print('Compute membership')
# t1 = time.clock()
if isUsingGPU == False:
Xl_cur = X_l[i]
Xu_cur = X_u[i]
classOfX = patClassId[i]
b = torch_memberG(Xl_cur, Xu_cur, self.V, self.W,
self.gamma)
else:
Xl_cur = X_l[i].cuda()
Xu_cur = X_u[i].cuda()
classOfX = patClassId[i].cuda()
b = gpu_memberG(Xl_cur, Xu_cur, self.V, self.W,
self.gamma)
# t2 = time.clock()
# print('Finish computing membership: ', t2 - t1)
# print(b)
bSort, index = torch.sort(b, descending=True)
# print('index = ', index)
# print('Self-class: ', self.classId)
# print(' i===', i)
if bSort[0] != 1 or (classOfX != self.classId[index[0]]
and classOfX != 0):
adjust = False
for j in index:
# test violation of max hyperbox size and class labels
if (classOfX == self.classId[j] or self.classId[j]
== 0 or classOfX == 0) and (
(torch.max(self.W[j], Xu_cur).float() -
torch.min(self.V[j], Xl_cur).float())
<= teta).all() == True:
# adjust the j-th hyperbox
self.V[j] = torch.min(self.V[j], Xl_cur)
self.W[j] = torch.max(self.W[j], Xu_cur)
indOfWinner = j
adjust = True
if classOfX != 0 and self.classId[j] == 0:
self.classId[j] = classOfX
if self.isDraw:
# Handle drawing graph
box_color = 'k'
if self.classId[j] < len(mark_col):
box_color = mark_col[self.classId[j]]
try:
listLines[j].remove()
except:
pass
if isUsingGPU == False:
hyperbox = drawbox(
np.asmatrix(self.V[
j,
0:np.minimum(xX, 3)].numpy()),
np.asmatrix(self.W[
j,
0:np.minimum(xX, 3)].numpy()),
drawing_canvas, box_color)
else:
hyperbox = drawbox(
np.asmatrix(
self.V[j, 0:np.minimum(xX, 3)].
cpu().numpy()),
np.asmatrix(
self.W[j, 0:np.minimum(xX, 3)].
cpu().numpy()), drawing_canvas,
box_color)
listLines[j] = hyperbox[0]
self.delay()
break
# if i-th sample did not fit into any existing box, create a new one
if not adjust:
# print('Create new hyperbox')
# t1 = time.clock()
self.V = torch.cat((self.V, Xl_cur.reshape(1, -1)),
0)
self.W = torch.cat((self.W, Xu_cur.reshape(1, -1)),
0)
if isUsingGPU == False:
self.classId = torch.cat(
(self.classId, torch.LongTensor([classOfX
])), 0)
else:
self.classId = torch.cat(
(self.classId,
torch.cuda.LongTensor([classOfX])), 0)
# t2 = time.clock()
# print('Finish compute new hyperbox: ', t2 - t1)
if self.isDraw:
# handle drawing graph
box_color = 'k'
if self.classId[-1] < len(mark_col):
box_color = mark_col[self.classId[-1]]
if isUsingGPU == False:
hyperbox = drawbox(
np.asmatrix(
X_l[i,
0:np.minimum(xX, 3)].numpy()),
np.asmatrix(
X_u[i,
0:np.minimum(xX, 3)].numpy()),
drawing_canvas, box_color)
else:
hyperbox = drawbox(
np.asmatrix(X_l[i,
0:np.minimum(xX, 3)].
cpu().numpy()),
np.asmatrix(X_u[i,
0:np.minimum(xX, 3)].
cpu().numpy()),
drawing_canvas, box_color)
listLines.append(hyperbox[0])
self.delay()
elif self.V.size(0) > 1:
for ii in range(self.V.size(0)):
if ii != indOfWinner and self.classId[
ii] != self.classId[indOfWinner]:
# print('Overlap Test')
# t1 = time.clock()
caseDim = torch_hyperboxOverlapTest(
self.V, self.W, indOfWinner,
ii) # overlap test
# t2 = time.clock()
# print('Finish overlap test: ', t2 - t1)
if len(caseDim) > 0:
# print('Hyperbox Contraction')
# t1 = time.clock()
self.V, self.W = torch_hyperboxContraction(
self.V, self.W, caseDim, ii,
indOfWinner)
# t2 = time.clock()
# print('Finish hyperbox Contaction: ', t2 - t1)
if self.isDraw:
# Handle graph drawing
boxii_color = boxwin_color = 'k'
if self.classId[ii] < len(
mark_col):
boxii_color = mark_col[
self.classId[ii]]
if self.classId[indOfWinner] < len(
mark_col):
boxwin_color = mark_col[
self.classId[indOfWinner]]
try:
listLines[ii].remove()
listLines[indOfWinner].remove()
except:
pass
if isUsingGPU == False:
hyperboxes = drawbox(
self.V[[ii, indOfWinner],
0:np.minimum(xX, 3)]
.numpy(),
self.W[[ii, indOfWinner],
0:np.minimum(xX, 3)]
.numpy(), drawing_canvas, [
boxii_color,
boxwin_color
])
else:
hyperboxes = drawbox(
self.V[[ii, indOfWinner],
0:np.minimum(xX, 3)]
.cpu().numpy(),
self.W[[ii, indOfWinner],
0:np.minimum(xX, 3)]
.cpu().numpy(),
drawing_canvas, [
boxii_color,
boxwin_color
])
listLines[ii] = hyperboxes[0]
listLines[
indOfWinner] = hyperboxes[1]
self.delay()
teta = teta * 0.9
if teta >= self.tMin:
result = torch_predict(self.V, self.W, self.classId, X_l, X_u,
patClassId, self.gamma, self.oper)
self.misclass = result.summis
# Draw last result
# if self.isDraw == True:
# # Handle drawing graph
# drawing_canvas.cla()
# color_ = np.empty(len(self.classId), dtype = object)
# for c in range(len(self.classId)):
# color_[c] = mark_col[self.classId[c]]
#
# drawbox(self.V[:, 0:np.minimum(xX, 3)], self.W[:, 0:np.minimum(xX, 3)], drawing_canvas, color_)
# self.delay()
#
# if self.isDraw:
# plt.show()
time_end = time.perf_counter()
self.elapsed_training_time = time_end - time_start
return self
def fit(self, X_l, X_u, patClassId):
"""
Xl Input data lower bounds (rows = objects, columns = features)
Xu Input data upper bounds (rows = objects, columns = features)
patClassId Input data class labels (crisp)
"""
if self.isNorm == True:
X_l, X_u = self.dataPreprocessing(X_l, X_u)
if isinstance(X_l, torch.Tensor) == False:
X_l = torch.from_numpy(X_l).float()
X_u = torch.from_numpy(X_u).float()
patClassId = torch.from_numpy(patClassId).long()
time_start = time.perf_counter()
isUsingGPU = False
if is_Have_GPU and X_l.size(0) * X_l.size(1) >= GPU_Computing_Threshold:
self.V = X_l.cuda()
self.W = X_u.cuda()
self.classId = patClassId.cuda()
isUsingGPU = True
else:
self.V = X_l
self.W = X_u
self.classId = patClassId
# training
isTraining = True
while isTraining:
isTraining = False
k = 0 # input pattern index
while k < len(self.classId):
idx_same_classes = (self.classId == self.classId[k]) | (self.classId == UNLABELED_CLASS) | ((self.classId != self.classId[k]) & (self.classId[k] == UNLABELED_CLASS))
idx_same_classes[k] = 0 # remove element in the position k
idex = torch.arange(len(self.classId))
idex = idex[idx_same_classes] # keep the indices of elements retained
V_same_class = self.V[idx_same_classes]
W_same_class = self.W[idx_same_classes]
if self.simil == 'short':
if isUsingGPU == False:
b = torch_memberG(self.W[k], self.V[k], V_same_class, W_same_class, self.gamma, self.oper, isUsingGPU)
else:
b = gpu_memberG(self.W[k], self.V[k], V_same_class, W_same_class, self.gamma, self.oper)
elif self.simil == 'long':
if isUsingGPU == False:
b = torch_memberG(self.V[k], self.W[k], V_same_class, W_same_class, self.gamma, self.oper, isUsingGPU)
else:
b = gpu_memberG(self.V[k], self.W[k], V_same_class, W_same_class, self.gamma, self.oper)
else:
b = torch_asym_similarity_one_many(self.V[k], self.W[k], V_same_class, W_same_class, self.gamma, self.sing, self.oper, isUsingGPU)
sortB, indB = torch.sort(b, descending=True)
idex = idex[indB]
maxB = sortB[sortB >= self.bthres] # apply membership threshold
if len(maxB) > 0:
idexmax = idex[sortB >= self.bthres]
if isUsingGPU == True:
idexmax = idexmax.cuda()
kMat = torch.cuda.LongTensor([k]).expand(idexmax.size(0))
else:
kMat = torch.LongTensor([k]).expand(idexmax.size(0))
pairewise_maxb = torch.cat((torch.min(kMat, idexmax).reshape(-1, 1).float(), torch.max(kMat,idexmax).reshape(-1, 1).float(), maxB.reshape(-1, 1)), dim=1)
if isUsingGPU:
els = torch.arange(pairewise_maxb.size(0)).cuda()
else:
els = torch.arange(pairewise_maxb.size(0))
for i in els:
# calculate new coordinates of k-th hyperbox by including pairewise_maxb(i,1)-th box, scrap the latter and leave the rest intact
# agglomorate pairewise_maxb(i, 0) and pairewise_maxb(i, 1) by adjusting pairewise_maxb(i, 0)
# remove pairewise_maxb(i, 1) by getting newV from 1 -> pairewise_maxb(i, 0) - 1, new coordinates for pairewise_maxb(i, 0), from pairewise_maxb(i, 0) + 1 -> pairewise_maxb(i, 1) - 1, pairewise_maxb(i, 1) + 1 -> end
row1 = pairewise_maxb[i, 0].long()
row2 = pairewise_maxb[i, 1].long()
newV = torch.cat((self.V[:row1], torch.min(self.V[row1], self.V[row2]).reshape(1, -1), self.V[row1 + 1:row2], self.V[row2 + 1:]), dim=0)
newW = torch.cat((self.W[:row1], torch.max(self.W[row1], self.W[row2]).reshape(1, -1), self.W[row1 + 1:row2], self.W[row2 + 1:]), dim=0)
newClassId = torch.cat((self.classId[:row2], self.classId[row2 + 1:]))
if (newClassId[row1] == UNLABELED_CLASS):
newClassId[row1] = self.classId[row2]
# adjust the hyperbox if no overlap and maximum hyperbox size is not violated
# position of adjustment is pairewise_maxb[i, 0] in new bounds
if ((((newW[row1] - newV[row1]) <= self.teta).all() == True) and (not torch_modifiedIsOverlap(newV, newW, row1, newClassId, isUsingGPU))):
self.V = newV
self.W = newW
self.classId = newClassId
isTraining = True
if k != row1: # position pairewise_maxb[i, 1] (also k) is removed, so next step should start from pairewise_maxb[i, 1]
k = k - 1
break # if hyperbox adjusted there's no need to look at other hyperboxes
k = k + 1
if isTraining == True and isUsingGPU == True and self.V.size(0) * self.V.size(1) < GPU_Computing_Threshold:
isUsingGPU = False
self.V = self.V.cpu()
self.W = self.W.cpu()
self.classId = self.classId.cpu()
time_end = time.perf_counter()
self.elapsed_training_time = time_end - time_start
return self
