def prediction(self, xtest, weighted=True, bestm=2, Y_prev=[]):
ypred = np.empty([np.shape(xtest)[0], self.ndim])
for n in range(0, np.shape(xtest)[0], 1):
w = np.empty([self.M, 1])
dw = np.empty([self.M, 1])
dc = np.empty([self.M, 1])
dcw = np.empty([self.M, 1])
yploc = np.empty([self.M, self.ndim])
var = np.empty([self.M, 1])
for k in range(0, self.M, 1):
try:
c = self.LocalData[k][2] #1x2
d = self.LocalData[k][3]
xW = np.dot((xtest[n] - c), self.W) # 1x2 X 2x2
w[k] = np.exp(-0.5 * np.dot(xW, np.transpose(
(xtest[n] - c))))
yploc[k], var[k] = self.Models[k].predict(xtest[n].reshape(
1, 2))
dw[k] = np.dot(
yploc[k] - Y_prev[-1].reshape(1, self.ndim),
np.transpose(yploc[k] -
Y_prev[-1].reshape(1, self.ndim)))
# dw[k] = np.linalg.norm(circstats.difference(yploc[k],Y_prev))
# dc[k] = np.dot(yploc[k]-d,np.transpose(yploc[k]-d))
# dcw[k] = np.dot(d-Y_prev,np.transpose(d-Y_prev))
except:
w[k] = 0
dw[k] = float("inf")
dc[k] = float("inf")
# dcw[k] = float("inf")
# var[k] = float("inf")
pass
if weighted:
if bestm > self.M:
h = self.M
else:
h = bestm
else:
h = 1
s = 10
wv = w * np.exp(-s * dw) / var
wv = np.nan_to_num(wv)
wv = wv.reshape(self.M, )
varmin = np.min(var) # minimum variance of local predictions
# wv = wv/np.sum(wv)
thresh = 0 # 0 uses all models
"Select for best models"
if np.max(wv) < thresh:
ind = wv == np.max(wv)
else:
ind = wv > thresh
"Weighted circular mean of predictions"
ypred[n] = circstats.mean(yploc, axis=0, w=wv.reshape(len(wv), 1))
"Normal Weighted mean"
# ypred[n] = np.dot(np.transpose(wv[ind]), yploc[ind]) / np.sum(wv[ind])
"Debug Prints"
# print("wv:" + str(wv))
# print("w:" + str(w))
# print("dw:" + str(dw))
# print("dc:" + str(dc))
# print("var:" + str(var))
# print("dcw:" + str(dcw))
# print("yploc:" + str(yploc))
return ypred, varmin
def prediction(self, xtest, weighted=True, bestm=2, Y_prev=[]):
ypred = np.empty([np.shape(xtest)[0], self.ndim])
# print(Y_prev)
# print(self.M)
for n in range(0, np.shape(xtest)[0], 1):
w = np.empty([self.M, 1])
dw = np.empty([self.M, 1])
dc = np.empty([self.M, 1])
dcw = np.empty([self.M, 1])
yploc = np.empty([self.M, self.ndim])
var = np.empty([self.M, 1])
for k in range(0, self.M, 1):
c = self.LocalData[k][2] #1x2
d = self.LocalData[k][3]
xW = np.dot((xtest[n] - c), self.W) # 1x2 X 22
# print(xW.shape)
# print(self.W.shape)
# print(xtest[n].shape)
# print(c.shape)
w[k] = np.exp(-0.5 * np.dot(xW, np.transpose((xtest[n] - c))))
# print(w[k])
yploc[k], var[k] = self.Models[k].predict(xtest[n].reshape(
1, 2))
dw[k] = np.dot(
yploc[k] - Y_prev[-1].reshape(1, self.ndim),
np.transpose(yploc[k] - Y_prev[-1].reshape(1, self.ndim)))
# try:
# # print(k)
# c = self.LocalData[k][2] #1x2
# d = self.LocalData[k][3]
# # d = self.UpdateY[k][-1]
# # dd = self.LocalData[k][1][-1]
# # dp = Y_prev
# # c = np.mean(self.Models[k].Z,axis = 0)
# # xW = np.dot((xtest[n]-c),self.W[k]) # 1x2 X 2x2
# # xW = self.Models[k].kern.K(xtest[n],c)
# xW = np.dot((xtest[n]-c),self.W) # 1x2 X 2x2
## print(xW)
# w[k] = np.exp(-0.5*np.dot(xW,np.transpose((xtest[n]-c))))
# print(w[k])
## w[k] = self.Models[k].kern.K(xtest[n].reshape(1,2),c.reshape(1,2))
# yploc[k], var[k] = self.Models[k].predict(xtest[n].reshape(1,2))
## print(yploc[k])
## Ypost = np.vstack((Y_prev,yploc[k]))
## print(Ypost)
## Ypost = np.unwrap(Ypost,axis=0)
## print(Ypost)
#
## yploc[k]= Ypost[-1].reshape(1,2)
## print(yploc[k])
## print(var[k])
## dw[k] = np.exp(-0.5*np.dot(yploc[k]-Y_prev,np.transpose(yploc[k]-Y_prev)))
# # dc[k] = np.exp(-0.5*np.dot(yploc[k]-d,np.transpose(yploc[k]-d)))
# # dcw[k] = np.exp(-0.5*np.dot(d-Y_prev,np.transpose(d-Y_prev)))
# dw[k] = np.dot(yploc[k]-Y_prev[-1].reshape(1,self.ndim),np.transpose(yploc[k]-Y_prev[-1].reshape(1,self.ndim)))
## dw[k] = np.linalg.norm(circstats.difference(yploc[k],Y_prev))
## dc[k] = np.dot(yploc[k]-d,np.transpose(yploc[k]-d))
## dcw[k] = np.dot(d-Y_prev,np.transpose(d-Y_prev))
# except:
# w[k] = 1
# dw[k] = float("inf")
# dc[k] = float("inf")
# pass
#
#
#
# yploc = yploc-2*np.pi
# yploc[yploc<0] += 2*np.pi
# yploc = np.unwrap(yploc)
# c = np.mean(self.mdrift.Z,axis = 0)
# xW = np.dot((xtest[n]-c),self.W) # 1x2 X 2x2
# w[k+1] = np.exp(-0.5*np.dot(xW,np.transpose((xtest[n]-c))))
# yploc[k+1], var[k+1] = self.mdrift.predict(xtest[n].reshape(1,2))
if weighted:
if bestm > self.M:
h = self.M
else:
h = bestm
else:
h = 1
# wv = np.multiply(w, np.exp(-var))
# wv = np.exp(-var)
# wv = w/var
wv = w * np.exp(-10 * dw) / var
# wv = w*dc*np.exp(-0.5*var)
# wv = w/(var*dc*dw)
# wv = w*np.exp(-10*dc)/var
# wv = w*np.exp(-dcw)/(var)
# wv = w/(var*(dcw+1e4))
wv = np.nan_to_num(wv)
# wv = dw/var
# wv = w/(dw*var)
# wv = w
# wv = w.reshape(self.M,1)
wv = wv.reshape(self.M, )
varmin = np.min(var)
# wv = w[w>0.9]
thresh = 0
# wv = wv/np.sum(wv)
if np.max(wv) < thresh:
# ind = np.argmax(wv)
ind = wv == np.max(wv)
else:
ind = wv > thresh
# ind = np.argmax(wv)
# print(np.argmax(wv))
# print(ind)
# print(wv.reshape(self.M,)>0.9)
# ind = np.argpartition(wv[:,0], -h)[-h:]
# print(ind)
# ypred[n] = np.dot(np.transpose(wv[wv>0.8]), yploc[wv>0.8]) / np.sum(wv[wv>0.8])
# ypred[n] = np.dot(np.transpose(wv[ind]), yploc[ind]) / np.sum(wv[ind])
# print(ypred)
# print(wv.shape)
# print(yploc.shape)
# print(yploc[0])
# yploc[yploc<0] += 2*np.pi
# yploc = np.unwrap(yploc)
# ypred[n] = self.circular_mean(wv[ind].reshape(np.sum(ind),1),yploc[ind])
# ypred[n] = circstats.mean(yploc[ind],axis = 0,w = wv[ind])
ypred[n] = circstats.mean(yploc, axis=0, w=wv.reshape(len(wv), 1))
# ypred[n] = np.dot(np.transpose(wv[ind]), yploc[ind]) / np.sum(wv[ind])
# ypred[n] = ypred[n] + 2*np.pi
# ypred[n][ypred[n]<0] += 2*np.pi
# ypred[n][ypred[n]>2*np.pi] -= 2*np.pi
# delta = ypred[n]-Y_prev
## new = Y_prev+delta*np.exp(-5*np.min(var))
# new = Y_prev+delta*np.max(wv)
# ypred[n] = new
# ypred[n] = np.array([self.circular_mean(wv,yploc[:,0]),self.circular_mean(wv,yploc[:,1])])
# print("wv:" + str(wv))
# print("w:" + str(w))
# print("dw:" + str(dw))
## print("dc:" + str(dc))
# print("var:" + str(var))
## print("dcw:" + str(dcw))
#
## print(ind)
#
# print(yploc[ind])
# print(ypred[n])
# print(dw)
return ypred, varmin
def prediction(self,xtest, weighted = True, bestm = 3, Y_prev = []):
ypred = np.empty([np.shape(xtest)[0], self.ndim])
for n in range(0, np.shape(xtest)[0], 1):
w = np.empty([self.M, 1])
dw = np.empty([self.M, 1])
dc = np.empty([self.M, 1])
dcw = np.empty([self.M, 1])
yploc = np.empty([self.M,self.ndim])
#xploc = np.empty([self.M,self.xdim])
var = np.empty([self.M,1])
for k in range(0, self.M, 1):
try:
c = self.LocalData[k][2] #1x2
d = self.LocalData[k][3]
xW = np.dot((xtest[n]-c),self.W) # 1x2 X 2x2
w[k] = np.exp(-0.5*np.dot(xW,np.transpose((xtest[n]-c))))
yploc[k], var[k] = self.Models[k].predict(xtest[n].reshape(1,self.xdim))
if Y_prev == []:
pass
else:
dcw[k] = np.dot(d-Y_prev[-1].reshape(1,self.ndim),np.transpose(d-Y_prev[-1].reshape(1,self.ndim)))
except:
w[k] = 0
dcw[k] = float("inf")
pass
if weighted:
if bestm > self.M:
h = self.M
else:
h = bestm
else:
h = 1
self.w = w
s = 0
if Y_prev == []:
wv = w/var
else:
wv = w*np.exp(-s*dcw)/var
wv =np.nan_to_num(wv)
wv = wv.reshape(self.M,)
varmin = np.min(var) # minimum variance of local predictions
thresh = 0 # 0 uses all models
"Select for best models"
# if np.max(wv) < thresh:
# ind = wv ==np.max(wv)
# else:
# ind = wv > thresh
ind = np.argpartition(wv, -h)[-h:]
# ypred[n] = np.dot(np.transpose(wv[ind]), yploc[ind]) / np.sum(wv[ind])
if self.encode:
"Normal Weighted mean"
ypred[n] = np.dot(np.transpose(wv[ind]), yploc[ind]) / np.sum(wv[ind])
else:
"Weighted circular mean of predictions"
ypred[n] = circstats.mean(yploc,axis = 0,w = wv.reshape(len(wv),1))
"Debug Prints"
#print("wv:" + str(wv))
#print("w:" + str(w))
# print("dw:" + str(dw))
# print("dc:" + str(dc))
#print("var:" + str(var))
#print("dcw:" + str(dcw))
#print("d:" + str(d))
#print("Yprev:" + str(Y_prev[-1].reshape(1,self.ndim)))
# print("yploc:" + str(yploc))
#print("xploc:" + str(xploc))
#print("ypred:" + str(ypred))
return ypred, varmin