def _eval_by_rel(questions, model, config, pred2idx, argw2idx,
relation_config):
embs1, ridxs, ys = [], [], []
all_cins = []
for q in questions:
x1 = utils.get_raw_event_repr(q.rel.e1, config, pred2idx, argw2idx)
embs1.append(x1)
ridxs.append(q.rel.rtype_idx)
y = q.ans_idx
ys.append(y)
cins = []
for c in q.choices:
cin = utils.get_raw_event_repr(c, config, pred2idx, argw2idx)
cins.append(cin.tolist())
all_cins.append(cins)
embs1 = torch.stack(embs1, dim=0).to(args.device)
ridxs = torch.LongTensor(ridxs).to(args.device)
ys = torch.LongTensor(ys).to(args.device)
embs1 = model._transfer(model.embed_event(embs1), ridxs)
rembs = model.rel_embeddings(ridxs)
all_cins = torch.LongTensor(all_cins).to(args.device)
n_choices = len(questions[0].choices)
scores = torch.zeros((n_choices, len(questions)), dtype=torch.float32)
for i in range(n_choices):
cembs = model.embed_event(all_cins[:, i, :])
_score = model._calc(embs1, model._transfer(cembs, ridxs), rembs)
score = torch.sum(_score, 1)
if model.norm > 1:
score = torch.pow(score, 1.0 / model.norm)
scores[i] = score
_, y_predict = torch.min(scores, 0)
return ys, y_predict
def build_examples(questions, config, pred2idx, argw2idx):
Xs, ys = [], []
for q in questions:
x1 = utils.get_raw_event_repr(q.rel.e1, config, pred2idx, argw2idx)
x2 = utils.get_raw_event_repr(q.rel.e2, config, pred2idx, argw2idx)
x = torch.cat((x1, x2), 0)
Xs.append(x)
y = q.rel.rtype_idx
ys.append(y)
Xs = torch.stack(Xs, dim=0).to(args.device)
ys = torch.LongTensor(ys).to(args.device)
return Xs, ys
def build_embeddings(model, questions, config, pred2idx, argw2idx, rtype, we):
e2idx = {}
idx = 0
for q in questions:
key = q.echain[0].__repr__()
if key not in e2idx:
e2idx[key] = idx
idx += 1
for clist in q.choice_lists:
for c in clist:
key = c.__repr__()
if key not in e2idx:
e2idx[key] = idx
idx += 1
e_len = 1 + config['arg0_max_len'] + config['arg1_max_len']
inputs = torch.zeros((len(e2idx), e_len),
dtype=torch.int64).to(args.device)
wdim = we[we.keys()[0]].shape[0]
w_embeddings = torch.zeros((len(e2idx), wdim),
dtype=torch.float32).to(args.device)
for q in questions:
key = q.echain[0].__repr__()
idx = e2idx[key]
inputs[idx] = utils.get_raw_event_repr(q.echain[0],
config,
pred2idx,
argw2idx,
device=args.device,
use_head=args.use_head)
w_embeddings[idx] = event_to_we(q.echain[0], we, wdim)
for clist in q.choice_lists:
for c in clist:
key = c.__repr__()
idx = e2idx[key]
inputs[idx] = utils.get_raw_event_repr(c,
config,
pred2idx,
argw2idx,
device=args.device,
use_head=args.use_head)
w_embeddings[idx] = event_to_we(c, we, wdim)
ev_embeddings = model._transfer(model.embed_event(inputs), rtype)
return e2idx, ev_embeddings, w_embeddings
def build_embeddings(model, questions, config, pred2idx, argw2idx, rtype):
e2idx = {}
idx = 0
for q in questions:
for ctx in q.get_contexts():
key = ctx.__repr__()
if key not in e2idx:
e2idx[key] = idx
idx += 1
for ch in q.choices:
key = ch.__repr__()
if key not in e2idx:
e2idx[key] = idx
idx += 1
e_len = 1 + config['arg0_max_len'] + config['arg1_max_len']
inputs = torch.zeros((len(e2idx), e_len),
dtype=torch.int64).to(args.device)
for q in questions:
for e in q.get_contexts():
idx = e2idx[e.__repr__()]
inputs[idx] = utils.get_raw_event_repr(e,
config,
pred2idx,
argw2idx,
device=args.device,
use_head=args.use_head)
for e in q.choices:
idx = e2idx[e.__repr__()]
inputs[idx] = utils.get_raw_event_repr(e,
config,
pred2idx,
argw2idx,
device=args.device,
use_head=args.use_head)
embeddings = model._transfer(model.embed_event(inputs), rtype)
return e2idx, embeddings
def build_ev_embeddings(questions, config, pred2idx, argw2idx, model):
with torch.no_grad():
idx = 0
e2idx = {}
xs = []
for i_cat in questions.keys():
for label, q in questions[i_cat]:
for e in [q.rel.e1, q.rel.e2]:
if e.__repr__() not in e2idx:
e2idx[e.__repr__()] = idx
idx += 1
x = utils.get_raw_event_repr(e, config, pred2idx,
argw2idx)
xs.append(x)
xs = torch.stack(xs, dim=0).to(args.device)
embeddings = model.embed_event(xs)
return e2idx, embeddings