def dry_run(self, mass):
batch = len(mass)
sss, labels, poss = handle_mass(mass)
pooled_embs = []
for ss, pos in zip(sss, poss):
left, right = get_left_right_by_ss_pos(ss, pos)
emb1 = B.compress_left_get_embs(self.bert, self.toker,
left) # (seq_len, 784)
emb2 = B.compress_right_get_embs(self.bert, self.toker,
right) # (seq_len, 784)
left, right = get_left_right_by_ss_pos(ss, pos, distant=2)
emb3 = B.compress_left_get_embs(self.bert, self.toker,
left) # (seq_len, 784)
emb4 = B.compress_right_get_embs(self.bert, self.toker,
right) # (seq_len, 784)
# print(f'{emb1.shape[0]}, {emb2.shape[0]}')
assert emb1.shape[0] == emb2.shape[0] == emb3.shape[
0] == emb4.shape[0]
mean = (emb1 + emb2 + emb3 + emb4) / 4 # (seq_len, 784)
pooled = mean.mean(0) # (784)
pooled_embs.append(pooled)
pooled_embs = t.stack(pooled_embs) # (batch, 784)
labels = t.LongTensor(labels) # (batch), (0 or 1)
if GPU_OK:
labels = labels.cuda()
o = self.classifier(pooled_embs) # (batch, 2)
self.print_train_info(o, labels, -1)
return o.argmax(1), labels
def train(self, mass):
batch = len(mass)
sss, labels, poss = handle_mass(mass)
pooled_embs = []
for ss, pos in zip(sss, poss):
left, right = get_left_right_by_ss_pos(ss, pos)
emb1 = B.compress_left_get_embs(self.bert, self.toker,
left) # (seq_len, 784)
emb2 = B.compress_right_get_embs(self.bert, self.toker,
right) # (seq_len, 784)
# print(f'{emb1.shape[0]}, {emb2.shape[0]}')
assert emb1.shape[0] == emb2.shape[0]
mean = (emb1 + emb2) / 2 # (seq_len, 784)
pooled = mean.mean(0) # (784)
pooled_embs.append(pooled)
pooled_embs = t.stack(pooled_embs) # (batch, 784)
labels = t.LongTensor(labels) # (batch), (0 or 1)
if GPU_OK:
labels = labels.cuda()
o = self.classifier(pooled_embs) # (batch, 1)
loss = self.cal_loss(o, labels)
self.zero_grad()
loss.backward()
self.optim.step()
self.print_train_info(o, labels, loss.detach().item())
return loss.detach().item()
def train(self, mass):
batch = len(mass)
sss, labels, poss = handle_mass(mass)
pooled_embs = []
clss_for_order_checking = [] # For order detector
order_labels = [] # For order detector
for ss, pos in zip(sss, poss):
left, right = get_left_right_by_ss_pos(ss, pos)
emb1 = B.compress_left_get_embs(self.bert, self.toker,
left) # (seq_len, 784)
emb2 = B.compress_right_get_embs(self.bert, self.toker,
right) # (seq_len, 784)
# print(f'{emb1.shape[0]}, {emb2.shape[0]}')
assert emb1.shape[0] == emb2.shape[0]
mean = (emb1 + emb2) / 2 # (seq_len, 784)
pooled = mean.mean(0) # (784)
pooled_embs.append(pooled)
# For order detector
if len(left) == 2:
if random.randrange(100) > 50: # 1/2的概率倒序
left_disturbed = list(reversed(left))
order_labels.append(1)
else:
left_disturbed = left.copy()
order_labels.append(0)
cls = B.compress_by_ss_pair_get_mean(self.bert, self.toker,
left_disturbed) # (784)
clss_for_order_checking.append(cls)
else:
print(f'Warning, left length = {len(left)}')
pooled_embs = t.stack(pooled_embs) # (batch, 784)
labels = t.LongTensor(labels) # (batch), (0 or 1)
order_labels = t.LongTensor(
order_labels) # (x <= batch) For order detector
if GPU_OK:
labels = labels.cuda()
order_labels = order_labels.cuda()
o = self.classifier(pooled_embs) # (batch, 1)
sector_loss = self.cal_loss(o, labels)
# For order detector
clss_for_order_checking = t.stack(
clss_for_order_checking) # (x <= batch, 784)
output_ordering = self.classifier2(clss_for_order_checking) # (x, 1)
ordering_loss = self.cal_loss(output_ordering, order_labels,
rate=0) # 不存在数据不均衡问题
loss = sector_loss + ordering_loss
self.zero_grad()
loss.backward()
self.optim.step()
self.print_train_info(o, labels, loss.detach().item())
return loss.detach().item()
def dry_run(self, mass):
batch = len(mass)
sss, labels, poss = handle_mass(mass)
pooled_embs = []
for ss, pos in zip(sss, poss):
left, right = get_left_right_by_ss_pos(ss, pos)
emb1 = B.compress_left_get_embs(self.bert, self.toker,
left) # (seq_len, 784)
emb2 = B.compress_right_get_embs(self.bert, self.toker,
right) # (seq_len, 784)
assert emb1.shape[0] == emb2.shape[0]
mean = (emb1 + emb2) / 2 # (seq_len, 784)
pooled = mean.mean(0) # (784)
pooled_embs.append(pooled)
pooled_embs = t.stack(pooled_embs) # (batch, 784)
labels = t.LongTensor(labels) # (batch), (0 or 1)
if GPU_OK:
labels = labels.cuda()
o = self.classifier(pooled_embs) # (batch, 1), sigmoided
self.print_train_info(o, labels, -1)
return fit_sigmoided_to_label(o), labels