def getpairs2(self, batch, params):
embed_size = self.memsize
Rel = self.getRel()
we = self.getWe()
# Rel0 = np.reshape(Rel,(-1,relsize))
newd = [
convertToIndex(i, self.words, we, self.rel, Rel) for i in batch
]
g1 = []
g2 = []
R = []
#print newd
length = len(batch)
for idx, e in enumerate(newd):
(r, t1, t2, s) = e
g1.append(t1)
g2.append(t2)
R.append(r)
#batch is list of tuples
p11 = []
p22 = []
p3 = []
if (params.type == 'MAX'):
for i in range(length):
#print 'i: ',i
id0 = R[i]
wpick = ['', '', '']
while (wpick[0] == ''):
index = random.randint(0, len(g1) - 1)
if (index != i):
wpick[0] = g1[index]
p11.append(wpick[0])
while (wpick[1] == ''):
index = random.randint(0, len(g2) - 1)
if (index != i):
wpick[1] = g2[index]
p22.append(wpick[1])
while (wpick[2] == ''):
index = random.randint(0, len(R) - 1)
if (index != i):
wpick[2] = R[index]
p3.append(wpick[2])
delim = (lookupwordID(we, self.words, "#"))
pT = [a + delim + b for a, b in zip(g1, g2)]
pTuple, pTupleMask = self.prepare_data(pT)
neT1 = [a + delim + b for a, b in zip(p11, g2)]
neTuple1, neTuple1Mask = self.prepare_data(neT1)
neT2 = [a + delim + b for a, b in zip(g1, p22)]
neTuple2, neTuple2Mask = self.prepare_data(neT2)
return (R, p3, pTuple, pTupleMask, neTuple1, neTuple1Mask, neTuple2,
neTuple2Mask)
def getpairs(self, batch, params):
relsize = self.relsize
Rel = self.getRel()
we = self.getWe()
Rel0 = np.reshape(Rel, (-1, relsize))
newd = [
convertToIndex(i, self.words, we, self.rel, Rel0) for i in batch
]
g1 = []
g2 = []
R = []
#print newd
length = len(batch)
#r0=np.zeros((length,relsize, relsize)).astype(theano.config.floatX)
#print relsize
for idx, e in enumerate(newd):
(r, t1, t2, s) = e
#print relsize
#print length
#print r
g1.append(t1)
g2.append(t2)
R.append(r)
#batch is list of tuples
g1x, g1mask = self.prepare_data(g1)
#maxlen = g1x.shape[1]
#print maxlen
g2x, g2mask = self.prepare_data(g2)
p1 = []
for i in range(length):
id0 = R[i]
min0 = -5000
p1.append(g1[random.randint(0, length - 1)])
p1x, p1mask = self.prepare_data(p1)
p2 = []
for i in range(length):
id0 = R[i]
min0 = -5000
p2.append(g2[random.randint(0, length - 1)])
p2x, p2mask = self.prepare_data(p2)
PR = []
for i in range(length):
id0 = R[i]
min0 = -5000
PR.append(R[random.randint(0, length - 1)])
return (g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x, p2mask, R, PR)
def getpairs(self, batch, params):
relsize = self.relsize
Rel = self.getRel()
we = self.getWe()
Rel0 = np.reshape(Rel, (-1, relsize))
newd = [
convertToIndex(i, self.words, we, self.rel, Rel0) for i in batch
]
g1 = []
g2 = []
R = []
length = len(batch)
#r0=np.zeros((length,relsize, relsize)).astype(theano.config.floatX)
for idx, e in enumerate(newd):
(r, t1, t2, s) = e
g1.append(t1)
g2.append(t2)
R.append(r)
#batch is list of tuples
g1x, g1mask, g1length = self.prepare_data(g1)
g2x, g2mask, g2length = self.prepare_data(g2)
embg1 = self.feedforward_function(g1x, g1mask, g1length)
embg2 = self.feedforward_function(g2x, g2mask, g2length)
p1 = []
p2 = []
neg_r = []
best = best1 = best2 = 1
if (params.type == 'MAX'):
for i in range(length):
id0 = R[i]
min0 = -5000
min1 = -5000
min2 = -5000
vec1 = embg1[i, :]
vec2 = embg2[i, :]
vec_r = Rel[id0, :, :]
for j in range(length):
if j != i:
gv1 = embg1[j, :]
temp1 = np.dot(gv1, vec_r)
np1 = np.inner(temp1, vec2)
if np1 > min0:
min0 = np1
best = j
for j1 in range(length):
if j1 != i:
gv2 = embg2[j1, :]
temp11 = np.dot(vec1, vec_r)
np11 = np.inner(temp1, gv2)
if np11 > min1:
min1 = np11
best1 = j1
for j2 in range(length):
if j2 != i:
id1 = R[j2]
matrix_r = Rel[id1, :, :]
temp111 = np.dot(vec1, matrix_r)
np111 = np.inner(temp111, vec2)
if np111 > min2:
min2 = np111
best2 = j2
# print best,best1,best2
p1.append(g1[best])
p2.append(g2[best1])
neg_r.append(R[best2])
if (params.type == 'MIX'):
for i in range(length):
r1 = randint(0, 1)
if r1 == 1:
id0 = R[i]
min0 = -5000
min1 = -5000
min2 = -5000
vec1 = embg1[i, :]
vec2 = embg2[i, :]
vec_r = Rel[id0, :, :]
for j in range(length):
if j != i:
gv1 = embg1[j, :]
temp1 = np.dot(gv1, vec_r)
np1 = np.inner(temp1, vec2)
if np1 > min0:
min0 = np1
best = j
for j1 in range(length):
if j1 != i:
gv2 = embg2[j1, :]
temp11 = np.dot(vec1, vec_r)
np11 = np.inner(temp1, gv2)
if np11 > min1:
min1 = np11
best1 = j1
for j2 in range(length):
if j2 != i:
id1 = R[j2]
matrix_r = Rel[id1, :, :]
temp111 = np.dot(vec1, matrix_r)
np111 = np.inner(temp111, vec2)
if np111 > min2:
min2 = np111
best2 = j2
p1.append(g1[best])
p2.append(g2[best1])
neg_r.append(R[best2])
else:
id0 = R[i]
wpick = ['', '', '']
while (wpick[0] == ''):
index = random.randint(0, len(g1) - 1)
if (index != i):
wpick[0] = g1[index]
p1.append(wpick[0])
while (wpick[1] == ''):
index = random.randint(0, len(g2) - 1)
if (index != i):
wpick[1] = g2[index]
p2.append(wpick[1])
while (wpick[2] == ''):
index = random.randint(0, len(R) - 1)
if (index != i):
wpick[2] = R[index]
neg_r.append(wpick[2])
if (params.type == 'RAND'):
for i in range(length):
id0 = R[i]
wpick = ['', '', '']
while (wpick[0] == ''):
index = random.randint(0, len(g1) - 1)
if (index != i):
wpick[0] = g1[index]
p1.append(wpick[0])
while (wpick[1] == ''):
index = random.randint(0, len(g2) - 1)
if (index != i):
wpick[1] = g2[index]
p2.append(wpick[1])
while (wpick[2] == ''):
index = random.randint(0, len(R) - 1)
if (index != i):
wpick[2] = R[index]
neg_r.append(wpick[2])
p1x, p1mask, p1length = self.prepare_data(p1)
p2x, p2mask, p2length = self.prepare_data(p2)
return (g1x, g1mask, g1length, g2x, g2mask, g2length, p1x, p1mask,
p1length, p2x, p2mask, p2length, R, neg_r)
def getpairs(self, batch, params):
relsize = self.relsize
Rel = self.getRel()
we = self.getWe()
Rel0 = np.reshape(Rel,(-1,relsize))
newd = [convertToIndex(i, self.words, we, self.rel, Rel0) for i in batch]
g1=[];g2=[];R=[]
#print newd
length = len(batch)
#r0=np.zeros((length,relsize, relsize)).astype(theano.config.floatX)
#print relsize
for idx, e in enumerate(newd):
(r, t1, t2, s) =e
#print relsize
#print length
#print r
g1.append(t1)
g2.append(t2)
R.append(r)
#batch is list of tuples
g1x, g1mask= self.prepare_data(g1)
#maxlen = g1x.shape[1]
#print maxlen
g2x, g2mask = self.prepare_data(g2)
embg1 = self.feedforward_function(g1x,g1mask)
embg2 = self.feedforward_function(g2x,g2mask)
p1=[]
for i in range(length):
id0 = R[i]
min0 = 5000
gv2=embg2[i,:]
vec_r = Rel[id0,:,:]
for j in range(length):
if j != i:
gv1=embg1[j,:]
#print vec_r
temp1 = np.dot(gv1, vec_r)
np1 = np.inner(temp1,gv2)
if np1 < min0:
min0=np1
best=j
if random.randint(0,1)!=0:
#print "get"
p1.append(g1[random.randint(0, length -1)])
else :
p1.append(g1[best])
p1x, p1mask = self.prepare_data(p1)
p2=[]
for i in range(length):
id0 = R[i]
min0 = 5000
gv1=embg1[i,:]
vec_r = Rel[id0,:,:]
for j in range(length):
if j != i:
gv2=embg2[j,:]
#print vec_r
temp1 = np.dot(gv1, vec_r)
np1 = np.inner(temp1,gv2)
if np1 < min0:
min0=np1
best=j
if random.randint(0,1)!=0:
#print "get"
p2.append(g2[random.randint(0, length -1)])
else :
p2.append(g2[best])
p2x, p2mask = self.prepare_data(p2)
PR=[]
for i in range(length):
id0 = R[i]
min0 = 5000
gv1=embg1[i,:]
gv2=embg2[i,:]
#vec_r = Rel[id0,:,:]
for j in range(length):
if R[j] != R[i]:
vr = Rel[R[j]]
#print vec_r
temp1 = np.dot(gv1, vr)
np1 = np.inner(temp1,gv2)
if np1 < min0:
min0=np1
best=j
if random.randint(0,1)!=0:
#print "get"
PR.append(R[random.randint(0, length -1)])
else :
PR.append(R[best])
return (g1x,g1mask,g2x,g2mask,p1x,p1mask, p2x, p2mask, R, PR)