def fix_dim(maxsize,
mindim,
word2ballDic=dict(),
bPath='/Users/tdong/data/glove/glove.6B/glove.6B.50Xball',
outputPath=""):
"""
:param maxsize:
:param mindim:
:param word2ballDic:
:param bPath:
:return:
"""
for bf in os.listdir(bPath):
with open(os.path.join(bPath, bf), 'r') as ifh:
wlst = ifh.readline().strip().split()
ballv = [decimal.Decimal(ele) for ele in wlst]
delta = maxsize - len(ballv)
if delta > 0:
assert len(wlst) < maxsize
print(bf, len(wlst), ballv[-1])
vec = vec_norm(ballv[:-2] +
[decimal.Decimal(mindim)] * delta) + ballv[-2:]
word2ballDic[bf] = vec
if outputPath:
create_ball_file(bf,
outputPath=bPath,
word2ballDic=word2ballDic)
return word2ballDic
def initialize_dictionaries(word2vecFile=None,
catDicFile=None,
wsChildrenFile=None):
"""
input is pre-trained word2vec
:param word2vecFile:
:param catDicFile:
:param wsChildrenFile:
:return:
"""
wscatCodeDic = dict()
wsChildrenDic = dict()
word2vecDic = dict()
if not os.path.isfile(word2vecFile):
print('file does not exist:', word2vecFile)
return
with open(word2vecFile, mode="r", encoding="utf-8") as w2v:
for line in w2v.readlines():
wlst = line.strip().split()
word2vecDic[wlst[0]] = vec_norm([float(ele) for ele in wlst[1:]])
if os.path.isfile(catDicFile):
with open(catDicFile, 'r') as cfh:
for ln in cfh.readlines():
wlst = ln[:-1].split()
wscatCodeDic[wlst[0]] = [int(ele) for ele in wlst[1:]]
if os.path.isfile(wsChildrenFile):
with open(wsChildrenFile, 'r') as chfh:
for ln in chfh:
wlst = ln[:-1].split()
wsChildrenDic[wlst[0]] = wlst[1:]
return wsChildrenDic, word2vecDic, wscatCodeDic
def initialize_ball(root,
addDim=[],
L0=0.1,
R0=0.1,
word2vecDic=dict(),
wscatCodeDic=dict(),
word2ballDic=dict(),
outputPath=None):
"""
:param root:
:param addDim:
:param L0:
:param R0:
:param word2vecDic:
:param wscatCodeDic:
:param word2ballDic:
:param outputPath:
:return:
"""
w2v = [
decimal.Decimal(ele * 100)
for ele in get_word2vector(root, word2vecDic=word2vecDic)
]
cpoint = w2v + [ele + 10 for ele in wscatCodeDic[root]] + addDim
word2ballDic[root] = vec_norm(cpoint) + [L0, R0]
if outputPath:
create_ball_file(root,
outputPath=outputPath,
word2ballDic=word2ballDic)
return word2ballDic[root], word2ballDic
def initialize_ball(root,
addDim=[],
L0=0.1,
R0=0.1,
word2vecDic=dict(),
wscatCodeDic=dict(),
word2ballDic=dict(),
outputPath=None):
"""
:param root:
:param addDim:
:param L0:
:param R0:
:param word2vecDic:
:param wscatCodeDic:
:param word2ballDic:
:param outputPath:
:return:
"""
w2v = [
decimal.Decimal(ele * 100)
for ele in get_word2vector(root, word2vecDic=word2vecDic)
]
# add catcode -> path from root to here e.g. 1 1 2 :> first tree, first child, second child
cpoint = w2v + [ele + 10 for ele in wscatCodeDic[root]] + addDim
word2ballDic[root] = vec_norm(cpoint) + [L0, R0]
print(f"Initialize Ball for root: {root}")
ball_generation_log.append(
Log(key=root,
operation=Operation.INITIALIZE,
operation_args=[],
vector=word2ballDic[root]))
if outputPath:
create_ball_file(root,
outputPath=outputPath,
word2ballDic=word2ballDic)
return word2ballDic[root], word2ballDic
def shift_whole_tree_of(tree,
deltaVec,
deltaL,
wsChildrenDic=dict(),
word2ballDic=dict(),
outputPath=None):
"""
:param tree:
:param deltaVec:
:param deltaL:
:param wsChildrenDic:
:param word2ballDic:
:param outputPath:
:return:
for child of tree:
shift_whole_tree_of(child, deltaVec, deltaL, outputPath=None)
l1, r1 = word2ballDic[tree][-2:]
l = np.sqrt(l1*l1 + deltaL*deltaL
+ 2*l1*deltaL* vec_cos(deltaVec, word2ballDic[tree][:-2]))
newVec = vec_norm(vec_point(word2ballDic[tree][:-2], l1) + vec_point(deltaVec, deltaL))
word2ballDic[tree] = list(newVec) + [l, r1]
for child of tree:
while True:
delta = qsr_DC_degree_by_name(child, tree)
if delta < 0:
word2ballDic[tree][-2] += - delta*1.01
else:
break
create_ball_file(tree, outputPath=outputPath)
"""
for child in get_children(tree,
wsChildrenDic=wsChildrenDic,
word2ballDic=word2ballDic):
word2ballDic = shift_whole_tree_of(child,
deltaVec,
deltaL,
wsChildrenDic=wsChildrenDic,
word2ballDic=word2ballDic,
outputPath=outputPath)
l1, r1 = word2ballDic[tree][-2:]
l = np.sqrt(l1 * l1 + deltaL * deltaL +
2 * l1 * deltaL * vec_cos(deltaVec, word2ballDic[tree][:-2]))
newVec = vec_norm(
vec_point(word2ballDic[tree][:-2], l1) + vec_point(deltaVec, deltaL))
word2ballDic[tree] = list(newVec) + [l, r1]
i, j, lst = 0, 0, get_children(tree,
wsChildrenDic=wsChildrenDic,
word2ballDic=word2ballDic)
for i in range(len(lst) - 1):
j = i + 1
while j < len(lst):
dcDelta = qsr_DC_degree(word2ballDic[lst[i]], word2ballDic[lst[j]])
if dcDelta < 0:
print(lst[j], lst[i], j, i)
word2ballDic = rotate_vector_till(lst[j],
lst[i],
word2ballDic=word2ballDic,
logFile='word2ball.log')
j += 1
for child in get_children(tree,
wsChildrenDic=wsChildrenDic,
word2ballDic=word2ballDic):
gap = 1
while True:
delta = qsr_P_degree(word2ballDic[child], word2ballDic[tree])
if delta < 0:
gap *= 2
word2ballDic[tree][-1] += -delta + gap
else:
break
if outputPath:
create_ball_file(tree,
outputPath=outputPath,
word2ballDic=word2ballDic)
return word2ballDic