def main(args):
init_property_of_dataset()
load_dataset()
print('relation size:', args.rel_size, 'entity size:', args.entity_size)
xp = Backend(args)
m = model.Model(args)
opt = get_opt(args)
opt.setup(m)
for epoch in range(args.epoch_size):
opt.alpha = args.beta0 / (1.0 + args.beta1 * epoch)
trLoss, Ntr = train(args, m, xp, opt)
trace('epoch:', epoch, 'tr Loss:', trLoss, Ntr)
dump_current_scores_of_devtest(args, m, xp)
def main(args):
loadflag = False
initilize_dataset()
args.rel_size, args.entity_size = get_sizes(args)
print('relation size:', args.rel_size, 'entity size:', args.entity_size)
RelCovVal = [1.0] * args.rel_size
RelCovValTmp = [0.0] * args.rel_size
EntCovVal = [1.0] * args.entity_size
EntCovValTmp = [0.0] * args.entity_size
xp = Backend(args)
m = get_model(args)
opt = get_opt(args)
if (loadflag == True):
serializers.load_npz(
'./savedWN/ModelA3CNN_OOKBtest_' + args.pooling_method + '.model',
m)
opt.setup(m)
for epoch in range(args.epoch_size):
'''
fileTmp = open('Ent100Eval_v4.txt','a')
for i in range(100):
fileTmp.write(str(EntCovVal[i])+'\t')
fileTmp.write('\n')
fileTmp.close()
'''
opt.alpha = args.beta0 / (1.0 + args.beta1 * epoch)
trLoss, Ntr = train(args, m, xp, opt, RelCovVal, RelCovValTmp,
EntCovVal, EntCovValTmp)
'''
for i in range(args.rel_size):
RelCovVal[i] = RelCovValTmp[i]
RelCovValTmp[i] = 0.0
for i in range(args.entity_size):
EntCovVal[i] = EntCovValTmp[i]
EntCovValTmp[i] = 0.0'''
tool.trace('epoch:', epoch, 'tr Loss:', tool.dress(trLoss), Ntr)
if (epoch % 5 == 0):
serializers.save_npz(
'./savedWN/ModelA3CNN_OOKBtest_' + args.pooling_method +
'.model', m)
dump_current_scores_of_devtest(args, m, xp, RelCovVal, EntCovVal)
def main(args):
loadflag = False
initilize_dataset()
args.rel_size,args.entity_size = get_sizes(args)
print('relation size:',args.rel_size,'entity size:',args.entity_size)
W1lists = []
file1 = open(args.pretrain_file + '/relation2vec.vec','r')
for line in file1.readlines():
line = line.strip().split('\t')
W1lists.append(line)
W1 = np.array(W1lists).astype(np.float32)
W2lists = []
file2 = open(args.pretrain_file + '/entity2vec.vec','r')
for line in file2.readlines():
line = line.strip().split('\t')
W2lists.append(line)
W2 = np.array(W2lists).astype(np.float32)
#print W1.shape, W2.shape
RelCovVal = [1.0] * args.rel_size
RelCovValTmp = [0.0] * args.rel_size
EntCovVal = [1.0] * args.entity_size
EntCovValTmp = [0.0] * args.entity_size
xp = Backend(args)
m = get_model(args,W1,W2)
opt = get_opt(args)
if(loadflag == True):
serializers.load_npz('./savedFB/ModelA3CNN_ATT_0817_head.model',m)
opt.setup(m)
normalizeR = m.Normal(xp)
#random.shuffle(train_data)
for epoch in range(args.epoch_size):
opt.alpha = args.beta0/(1.0+args.beta1*epoch)
trLoss,Ntr = train(args,m,xp,opt,RelCovVal,RelCovValTmp,EntCovVal,EntCovValTmp)
normalizeR = m.Normal(xp)
tool.trace('epoch:',epoch,'tr Loss:',tool.dress(trLoss),Ntr)
if(epoch % 5 == 0):
serializers.save_npz('./savedFB/ModelA3CNN_ATT_0817_head.model',m)
dump_current_scores_of_devtest(args,m,xp,RelCovVal,EntCovVal)
def main(args):
# 读入规则
rules_addition()
# 初始化所有数据集
initilize_dataset()
# 根据文件内容修改关系和实体的默认值
args.rel_size, args.entity_size = get_sizes(args)
print('relation size:', args.rel_size, 'entity size:', args.entity_size)
xp = Backend(args) #返回一个可调用的对象
m = get_model(args) # return A0(args)
# Setup an optimizer
# 设置训练时用的优化方法，默认为Adam
opt = get_opt(args) # return optimizers.Adam()
# setup()方法只是为优化器提供一个link
opt.setup(m)
for epoch in range(args.epoch_size):
opt.alpha = args.beta0 / (1.0 + args.beta1 * epoch)
trLoss, Ntr = train(args, m, xp, opt) #对应到main.py的train
tool.trace('epoch:', epoch, 'tr Loss:', tool.dress(trLoss), Ntr)
dump_current_scores_of_devtest(args, m, xp)
def main(args):
loadflag = False
initilize_dataset()
args.rel_size, args.entity_size = get_sizes(args)
print('relation size:', args.rel_size, 'entity size:', args.entity_size)
xp = Backend(args)
m = get_model(args)
opt = get_opt(args)
if (loadflag == True):
serializers.load_npz(
'./savedWN/ModelA0_OOKBtest_' + args.pooling_method, m)
opt.setup(m)
for epoch in range(args.epoch_size):
#dump_current_scores_of_devtest(args,m,xp)
opt.alpha = args.beta0 / (1.0 + args.beta1 * epoch)
trLoss, Ntr = train(args, m, xp, opt)
tool.trace('epoch:', epoch, 'tr Loss:', tool.dress(trLoss), Ntr)
if (epoch % 10 == 0):
serializers.save_npz(
'./savedWN/ModelA0_OOKBtest_' + args.pooling_method, m)
dump_current_scores_of_devtest(args, m, xp)
def main(args):
global candidate_heads, gold_heads, candidate_tails, gold_tails, black_set
xp = XP(args)
args.rel_size, args.entity_size = get_sizes(args)
print('relation size:', args.rel_size, 'entity size:', args.entity_size)
m = get_model(args)
opt = get_opt(args)
opt.setup(m)
relations = dict()
links = defaultdict(set)
for line in tool.read(args.train_file):
items = list(map(int, line.strip().split('\t')))
if len(items) == 4:
h, r, t, l = items
if l == 0: continue
else:
h, r, t = items
relations[(h, t)] = r
links[t].add(h)
links[h].add(t)
gold_heads[(r, t)].add(h)
gold_tails[(h, r)].add(t)
candidate_heads[r].add(h)
candidate_tails[r].add(t)
tail_per_head[h].add(t)
head_per_tail[t].add(h)
for e in links:
links[e] = list(links[e])
for p in gold_heads:
if len(candidate_heads[p[0]] - gold_heads[p]) == 0:
p = (-p[0], p[1])
black_set.add(p)
for p in gold_tails:
if len(candidate_tails[p[1]] - gold_tails[p]) == 0:
black_set.add(p)
print('black list size:', len(black_set))
for r in candidate_heads:
candidate_heads[r] = list(candidate_heads[r])
for r in candidate_tails:
candidate_tails[r] = list(candidate_tails[r])
for h in tail_per_head:
tail_per_head[h] = len(tail_per_head[h]) + 0.0
for t in head_per_tail:
head_per_tail[t] = len(head_per_tail[t]) + 0.0
if args.train_file == args.auxiliary_file:
tool.trace('use: edges=links')
edges = links
else:
tool.trace('use: different edges')
edges = defaultdict(set)
for line in tool.read(args.auxiliary_file):
items = list(map(int, line.strip().split('\t')))
if len(items) == 4:
h, r, t, l = items
if l == 0: continue
else:
h, r, t = items
relations[(h, t)] = r
edges[t].add(h)
edges[h].add(t)
for e in edges:
edges[e] = list(edges[e])
for epoch in range(args.epoch_size):
opt.alpha = args.beta0 / (1.0 + args.beta1 * epoch)
trLoss, Ntr = train(args, m, xp, opt, links, relations, edges)
evaluate(args, m, xp, links, relations, edges)
tool.trace('epoch:', epoch, 'tr Loss:', tool.dress(trLoss), Ntr)