原文:http://nlp.seas.harvard.edu/2018/04/03/attention.html
作者:Alexander Rush
译者:哈工大SCIR 刘元兴
本文转载自:哈工大SCIR
请先阅读Transformer注解及PyTorch实现(上)
训练
- 批和掩码
- 训练循环
- 训练数据和批处理
- 硬件和训练进度
- 优化器
- 正则化
- 标签平滑
第一个例子
- 数据生成
- 损失计算
- 贪心解码
真实示例
- 数据加载
- 迭代器
- 多GPU训练
- 训练系统附加组件:BPE,搜索,平均
结果
- 注意力可视化
结论
训练
本节介绍模型的训练方法。
快速穿插介绍训练标准编码器解码器模型需要的一些工具。首先我们定义一个包含源和目标句子的批训练对象用于训练,同时构造掩码。
class Batch:
"Object for holding a batch of data with mask during training."
def __init__(self, src, trg=None, pad=0):
self.src = src
self.src_mask = (src != pad).unsqueeze(-2)
if trg is not None:
self.trg = trg[:, :-1]
self.trg_y = trg[:, 1:]
self.trg_mask = \
self.make_std_mask(self.trg, pad)
self.ntokens = (self.trg_y != pad).data.sum()
@staticmethod
def make_std_mask(tgt, pad):
"Create a mask to hide padding and future words."
tgt_mask = (tgt != pad).unsqueeze(-2)
tgt_mask = tgt_mask & Variable(
subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data))
return tgt_mask
接下来,我们创建一个通用的训练和得分函数来跟踪损失。我们传入一个通用的损失计算函数,它也处理参数更新。
def run_epoch(data_iter, model, loss_compute):
"Standard Training and Logging Function"
start = time.time()
total_tokens = 0
total_loss = 0
tokens = 0
for i, batch in enumerate(data_iter):
out = model.forward(batch.src, batch.trg,
batch.src_mask, batch.trg_mask)
loss = loss_compute(out, batch.trg_y, batch.ntokens)
total_loss += loss
total_tokens += batch.ntokens
tokens += batch.ntokens
if i % 50 == 1:
elapsed = time.time() - start
print("Epoch Step: %d Loss: %f Tokens per Sec: %f" %
(i, loss / batch.ntokens, tokens / elapsed))
start = time.time()
tokens = 0
return total_loss / total_tokens
我们使用标准WMT 2014英语-德语数据集进行了训练,该数据集包含大约450万个句子对。使用字节对的编码方法对句子进行编码,该编码具有大约37000个词的共享源-目标词汇表。对于英语-法语,我们使用了WMT 2014 英语-法语数据集,该数据集由36M个句子组成,并将词分成32000个词片(Word-piece)的词汇表。
句子对按照近似的序列长度进行批处理。每个训练批包含一组句子对,包含大约25000个源词和25000个目标词。
我们将使用torch text来创建批次。下面更详细地讨论实现过程。我们在torchtext的一个函数中创建批次,确保填充到最大批训练长度的大小不超过阈值(如果我们有8个GPU,则阈值为25000)。
global max_src_in_batch, max_tgt_in_batch
def batch_size_fn(new, count, sofar):
"Keep augmenting batch and calculate total number of tokens + padding."
global max_src_in_batch, max_tgt_in_batch
if count == 1:
max_src_in_batch = 0
max_tgt_in_batch = 0
max_src_in_batch = max(max_src_in_batch, len(new.src))
max_tgt_in_batch = max(max_tgt_in_batch, len(new.trg) + 2)
src_elements = count * max_src_in_batch
tgt_elements = count * max_tgt_in_batch
return max(src_elements, tgt_elements)
我们在一台配备8个NVIDIA P100 GPU的机器上训练我们的模型。对于使用本文所述的超参数的基本模型,每个训练单步大约需要0.4秒。我们对基础模型进行了总共100,000步或12小时的训练。对于我们的大型模型,每个训练单步时间为1.0秒。大型模型通常需要训练300,000步(3.5天)。
我们选择Adam[1]作为优化器 ,其参数为、和。根据以下公式,我们在训练过程中改变了学习率:。在预热中随步数线性地增加学习速率,并且此后与步数的反平方根成比例地减小它。我们设置预热步数为4000。
注意:这部分非常重要,需要这种设置训练模型。
class NoamOpt:
"Optim wrapper that implements rate."
def __init__(self, model_size, factor, warmup, optimizer):
self.optimizer = optimizer
self._step = 0
self.warmup = warmup
self.factor = factor
self.model_size = model_size
self._rate = 0
def step(self):
"Update parameters and rate"
self._step += 1
rate = self.rate()
for p in self.optimizer.param_groups:
p['lr'] = rate
self._rate = rate
self.optimizer.step()
def rate(self, step = None):
"Implement `lrate` above"
if step is None:
step = self._step
return self.factor * \
(self.model_size ** (-0.5) *
min(step ** (-0.5), step * self.warmup ** (-1.5)))
def get_std_opt(model):
return NoamOpt(model.src_embed[0].d_model, 2, 4000,
torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))
当前模型在不同模型大小和超参数的情况下的曲线示例。
# Three settings of the lrate hyperparameters.
opts = [NoamOpt(512, 1, 4000, None),
NoamOpt(512, 1, 8000, None),
NoamOpt(256, 1, 4000, None)]
plt.plot(np.arange(1, 20000), [[opt.rate(i) for opt in opts] for i in range(1, 20000)])
plt.legend(["512:4000", "512:8000", "256:4000"])
None
正则化
在训练期间,我们采用了值 [2]的标签平滑。这种做法提高了困惑度,因为模型变得更加不确定,但提高了准确性和BLEU分数。
我们使用KL div loss实现标签平滑。相比使用独热目标分布,我们创建一个分布,其包含正确单词的置信度和整个词汇表中分布的其余平滑项。
class LabelSmoothing(nn.Module):
"Implement label smoothing."
def __init__(self, size, padding_idx, smoothing=0.0):
super(LabelSmoothing, self).__init__()
self.criterion = nn.KLDivLoss(size_average=False)
self.padding_idx = padding_idx
self.confidence = 1.0 - smoothing
self.smoothing = smoothing
self.size = size
self.true_dist = None
def forward(self, x, target):
assert x.size(1) == self.size
true_dist = x.data.clone()
true_dist.fill_(self.smoothing / (self.size - 2))
true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
true_dist[:, self.padding_idx] = 0
mask = torch.nonzero(target.data == self.padding_idx)
if mask.dim() > 0:
true_dist.index_fill_(0, mask.squeeze(), 0.0)
self.true_dist = true_dist
return self.criterion(x, Variable(true_dist, requires_grad=False))
在这里,我们可以看到标签平滑的示例。
# Example of label smoothing.
crit = LabelSmoothing(5, 0, 0.4)
predict = torch.FloatTensor([[0, 0.2, 0.7, 0.1, 0],
[0, 0.2, 0.7, 0.1, 0],
[0, 0.2, 0.7, 0.1, 0]])
v = crit(Variable(predict.log()),
Variable(torch.LongTensor([2, 1, 0])))
# Show the target distributions expected by the system.
plt.imshow(crit.true_dist)
None
如果对给定的选择非常有信心,标签平滑实际上会开始惩罚模型。
crit = LabelSmoothing(5, 0, 0.1)
def loss(x):
d = x + 3 * 1
predict = torch.FloatTensor([[0, x / d, 1 / d, 1 / d, 1 / d],
])
#print(predict)
return crit(Variable(predict.log()),
Variable(torch.LongTensor([1]))).data[0]
plt.plot(np.arange(1, 100), [loss(x) for x in range(1, 100)])
None
我们可以先尝试一个简单的复制任务。给定来自小词汇表的随机输入符号集,目标是生成那些相同的符号。
def data_gen(V, batch, nbatches):
"Generate random data for a src-tgt copy task."
for i in range(nbatches):
data = torch.from_numpy(np.random.randint(1, V, size=(batch, 10)))
data[:, 0] = 1
src = Variable(data, requires_grad=False)
tgt = Variable(data, requires_grad=False)
yield Batch(src, tgt, 0)
class SimpleLossCompute:
"A simple loss compute and train function."
def __init__(self, generator, criterion, opt=None):
self.generator = generator
self.criterion = criterion
self.opt = opt
def __call__(self, x, y, norm):
x = self.generator(x)
loss = self.criterion(x.contiguous().view(-1, x.size(-1)),
y.contiguous().view(-1)) / norm
loss.backward()
if self.opt is not None:
self.opt.step()
self.opt.optimizer.zero_grad()
return loss.data[0] * norm
# Train the simple copy task.
V = 11
criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
model = make_model(V, V, N=2)
model_opt = NoamOpt(model.src_embed[0].d_model, 1, 400,
torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))
for epoch in range(10):
model.train()
run_epoch(data_gen(V, 30, 20), model,
SimpleLossCompute(model.generator, criterion, model_opt))
model.eval()
print(run_epoch(data_gen(V, 30, 5), model,
SimpleLossCompute(model.generator, criterion, None)))
Epoch Step: 1 Loss: 3.023465 Tokens per Sec: 403.074173
Epoch Step: 1 Loss: 1.920030 Tokens per Sec: 641.689380
1.9274832487106324
Epoch Step: 1 Loss: 1.940011 Tokens per Sec: 432.003378
Epoch Step: 1 Loss: 1.699767 Tokens per Sec: 641.979665
1.657595729827881
Epoch Step: 1 Loss: 1.860276 Tokens per Sec: 433.320240
Epoch Step: 1 Loss: 1.546011 Tokens per Sec: 640.537198
1.4888023376464843
Epoch Step: 1 Loss: 1.682198 Tokens per Sec: 432.092305
Epoch Step: 1 Loss: 1.313169 Tokens per Sec: 639.441857
1.3485562801361084
Epoch Step: 1 Loss: 1.278768 Tokens per Sec: 433.568756
Epoch Step: 1 Loss: 1.062384 Tokens per Sec: 642.542067
0.9853351473808288
Epoch Step: 1 Loss: 1.269471 Tokens per Sec: 433.388727
Epoch Step: 1 Loss: 0.590709 Tokens per Sec: 642.862135
0.5686767101287842
Epoch Step: 1 Loss: 0.997076 Tokens per Sec: 433.009746
Epoch Step: 1 Loss: 0.343118 Tokens per Sec: 642.288427
0.34273059368133546
Epoch Step: 1 Loss: 0.459483 Tokens per Sec: 434.594030
Epoch Step: 1 Loss: 0.290385 Tokens per Sec: 642.519464
0.2612409472465515
Epoch Step: 1 Loss: 1.031042 Tokens per Sec: 434.557008
Epoch Step: 1 Loss: 0.437069 Tokens per Sec: 643.630322
0.4323212027549744
Epoch Step: 1 Loss: 0.617165 Tokens per Sec: 436.652626
Epoch Step: 1 Loss: 0.258793 Tokens per Sec: 644.372296
0.27331129014492034
为简单起见,此代码使用贪心解码来预测翻译。
def greedy_decode(model, src, src_mask, max_len, start_symbol):
memory = model.encode(src, src_mask)
ys = torch.ones(1, 1).fill_(start_symbol).type_as(src.data)
for i in range(max_len-1):
out = model.decode(memory, src_mask,
Variable(ys),
Variable(subsequent_mask(ys.size(1))
.type_as(src.data)))
prob = model.generator(out[:, -1])
_, next_word = torch.max(prob, dim = 1)
next_word = next_word.data[0]
ys = torch.cat([ys,
torch.ones(1, 1).type_as(src.data).fill_(next_word)], dim=1)
return ys
model.eval()
src = Variable(torch.LongTensor([[1,2,3,4,5,6,7,8,9,10]]) )
src_mask = Variable(torch.ones(1, 1, 10) )
print(greedy_decode(model, src, src_mask, max_len=10, start_symbol=1))
1 2 3 4 5 6 7 8 9 10
[torch.LongTensor of size 1x10]
现在我们通过IWSLT德语-英语翻译任务介绍一个真实示例。该任务比上文提及的WMT任务小得多,但它说明了整个系统。我们还展示了如何使用多个GPU处理加速其训练。
#!pip install torchtext spacy
#!python -m spacy download en
#!python -m spacy download de
我们将使用torchtext和spacy加载数据集以进行词语切分。
# For data loading.
from torchtext import data, datasets
if True:
import spacy
spacy_de = spacy.load('de')
spacy_en = spacy.load('en')
def tokenize_de(text):
return [tok.text for tok in spacy_de.tokenizer(text)]
def tokenize_en(text):
return [tok.text for tok in spacy_en.tokenizer(text)]
BOS_WORD = '<s>'
EOS_WORD = '</s>'
BLANK_WORD = "<blank>"
SRC = data.Field(tokenize=tokenize_de, pad_token=BLANK_WORD)
TGT = data.Field(tokenize=tokenize_en, init_token = BOS_WORD,
eos_token = EOS_WORD, pad_token=BLANK_WORD)
MAX_LEN = 100
train, val, test = datasets.IWSLT.splits(
exts=('.de', '.en'), fields=(SRC, TGT),
filter_pred=lambda x: len(vars(x)['src']) <= MAX_LEN and
len(vars(x)['trg']) <= MAX_LEN)
MIN_FREQ = 2
SRC.build_vocab(train.src, min_freq=MIN_FREQ)
TGT.build_vocab(train.trg, min_freq=MIN_FREQ)
批训练对于速度来说很重要。我们希望批次分割非常均匀并且填充最少。要做到这一点,我们必须修改torchtext默认的批处理函数。这部分代码修补其默认批处理函数,以确保我们搜索足够多的句子以构建紧密批处理。
class MyIterator(data.Iterator):
def create_batches(self):
if self.train:
def pool(d, random_shuffler):
for p in data.batch(d, self.batch_size * 100):
p_batch = data.batch(
sorted(p, key=self.sort_key),
self.batch_size, self.batch_size_fn)
for b in random_shuffler(list(p_batch)):
yield b
self.batches = pool(self.data(), self.random_shuffler)
else:
self.batches = []
for b in data.batch(self.data(), self.batch_size,
self.batch_size_fn):
self.batches.append(sorted(b, key=self.sort_key))
def rebatch(pad_idx, batch):
"Fix order in torchtext to match ours"
src, trg = batch.src.transpose(0, 1), batch.trg.transpose(0, 1)
return Batch(src, trg, pad_idx)
最后为了真正地快速训练,我们将使用多个GPU。这部分代码实现了多GPU字生成。它不是Transformer特有的,所以我不会详细介绍。其思想是将训练时的单词生成分成块,以便在许多不同的GPU上并行处理。我们使用PyTorch并行原语来做到这一点:
复制 - 将模块拆分到不同的GPU上
分散 - 将批次拆分到不同的GPU上
并行应用 - 在不同GPU上将模块应用于批处理
聚集 - 将分散的数据聚集到一个GPU上
nn.DataParallel - 一个特殊的模块包装器,在评估之前调用它们。
# Skip if not interested in multigpu.
class MultiGPULossCompute:
"A multi-gpu loss compute and train function."
def __init__(self, generator, criterion, devices, opt=None, chunk_size=5):
# Send out to different gpus.
self.generator = generator
self.criterion = nn.parallel.replicate(criterion,
devices=devices)
self.opt = opt
self.devices = devices
self.chunk_size = chunk_size
def __call__(self, out, targets, normalize):
total = 0.0
generator = nn.parallel.replicate(self.generator,
devices=self.devices)
out_scatter = nn.parallel.scatter(out,
target_gpus=self.devices)
out_grad = [[] for _ in out_scatter]
targets = nn.parallel.scatter(targets,
target_gpus=self.devices)
# Divide generating into chunks.
chunk_size = self.chunk_size
for i in range(0, out_scatter[0].size(1), chunk_size):
# Predict distributions
out_column = [[Variable(o[:, i:i+chunk_size].data,
requires_grad=self.opt is not None)]
for o in out_scatter]
gen = nn.parallel.parallel_apply(generator, out_column)
# Compute loss.
y = [(g.contiguous().view(-1, g.size(-1)),
t[:, i:i+chunk_size].contiguous().view(-1))
for g, t in zip(gen, targets)]
loss = nn.parallel.parallel_apply(self.criterion, y)
# Sum and normalize loss
l = nn.parallel.gather(loss,
target_device=self.devices[0])
l = l.sum()[0] / normalize
total += l.data[0]
# Backprop loss to output of transformer
if self.opt is not None:
l.backward()
for j, l in enumerate(loss):
out_grad[j].append(out_column[j][0].grad.data.clone())
# Backprop all loss through transformer.
if self.opt is not None:
out_grad = [Variable(torch.cat(og, dim=1)) for og in out_grad]
o1 = out
o2 = nn.parallel.gather(out_grad,
target_device=self.devices[0])
o1.backward(gradient=o2)
self.opt.step()
self.opt.optimizer.zero_grad()
return total * normalize
现在我们创建模型,损失函数,优化器,数据迭代器和并行化。
# GPUs to use
devices = [0, 1, 2, 3]
if True:
pad_idx = TGT.vocab.stoi["<blank>"]
model = make_model(len(SRC.vocab), len(TGT.vocab), N=6)
model.cuda()
criterion = LabelSmoothing(size=len(TGT.vocab), padding_idx=pad_idx, smoothing=0.1)
criterion.cuda()
BATCH_SIZE = 12000
train_iter = MyIterator(train, batch_size=BATCH_SIZE, device=0,
repeat=False, sort_key=lambda x: (len(x.src), len(x.trg)),
batch_size_fn=batch_size_fn, train=True)
valid_iter = MyIterator(val, batch_size=BATCH_SIZE, device=0,
repeat=False, sort_key=lambda x: (len(x.src), len(x.trg)),
batch_size_fn=batch_size_fn, train=False)
model_par = nn.DataParallel(model, device_ids=devices)
None
现在我们训练模型。我将稍微使用预热步骤,但其他一切都使用默认参数。在具有4个Tesla V100 GPU的AWS p3.8xlarge机器上,每秒运行约27,000个词,批训练大小大小为12,000。
#!wget https://s3.amazonaws.com/opennmt-models/iwslt.pt
if False:
model_opt = NoamOpt(model.src_embed[0].d_model, 1, 2000,
torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))
for epoch in range(10):
model_par.train()
run_epoch((rebatch(pad_idx, b) for b in train_iter),
model_par,
MultiGPULossCompute(model.generator, criterion,
devices=devices, opt=model_opt))
model_par.eval()
loss = run_epoch((rebatch(pad_idx, b) for b in valid_iter),
model_par,
MultiGPULossCompute(model.generator, criterion,
devices=devices, opt=None))
print(loss)
else:
model = torch.load("iwslt.pt")
一旦训练完成,我们可以解码模型以产生一组翻译。在这里,我们只需翻译验证集中的第一个句子。此数据集非常小,因此使用贪婪搜索的翻译相当准确。
for i, batch in enumerate(valid_iter):
src = batch.src.transpose(0, 1)[:1]
src_mask = (src != SRC.vocab.stoi["<blank>"]).unsqueeze(-2)
out = greedy_decode(model, src, src_mask,
max_len=60, start_symbol=TGT.vocab.stoi["<s>"])
print("Translation:", end="\t")
for i in range(1, out.size(1)):
sym = TGT.vocab.itos[out[0, i]]
if sym == "</s>": break
print(sym, end =" ")
print()
print("Target:", end="\t")
for i in range(1, batch.trg.size(0)):
sym = TGT.vocab.itos[batch.trg.data[i, 0]]
if sym == "</s>": break
print(sym, end =" ")
print()
break
Translation: <unk> <unk> . In my language , that means , thank you very much .
Gold: <unk> <unk> . It means in my language , thank you very much .
所以这主要涵盖了Transformer模型本身。有四个方面我们没有明确涵盖。我们还实现了所有这些附加功能 OpenNMT-py[3].
1) 字节对编码/ 字片(Word-piece):我们可以使用库来首先将数据预处理为子字单元。参见Rico Sennrich的subword-nmt实现[4]。这些模型将训练数据转换为如下所示:
▁Die ▁Protokoll datei ▁kann ▁ heimlich ▁per ▁E - Mail ▁oder ▁FTP ▁an ▁einen ▁bestimmte n ▁Empfänger ▁gesendet ▁werden .
2) 共享嵌入:当使用具有共享词汇表的BPE时,我们可以在源/目标/生成器之间共享相同的权重向量,详细见[5]。要将其添加到模型,只需执行以下操作:
if False:
model.src_embed[0].lut.weight = model.tgt_embeddings[0].lut.weight
model.generator.lut.weight = model.tgt_embed[0].lut.weight
3) 集束搜索:这里展开说有点太复杂了。PyTorch版本的实现可以参考 OpenNMT- py[6]。
4) 模型平均:这篇文章平均最后k个检查点以创建一个集合效果。如果我们有一堆模型,我们可以在事后这样做:
def average(model, models):
"Average models into model"
for ps in zip(*[m.params() for m in [model] + models]):
p[0].copy_(torch.sum(*ps[1:]) / len(ps[1:]))
在WMT 2014英语-德语翻译任务中,大型Transformer模型(表2中的Transformer(大))优于先前报告的最佳模型(包括集成的模型)超过2.0 BLEU,建立了一个新的最先进BLEU得分为28.4。该模型的配置列于表3的底部。在8个P100 GPU的机器上,训练需要需要3.5天。甚至我们的基础模型也超过了之前发布的所有模型和集成,而且只占培训成本的一小部分。
在WMT 2014英语-法语翻译任务中,我们的大型模型获得了41.0的BLEU分数,优于以前发布的所有单一模型,不到以前最先进技术培训成本的1/4 模型。使用英语到法语训练的Transformer(大)模型使用dropout概率 = 0.1,而不是0.3。
Image(filename="images/results.png")
我们在这里编写的代码是基本模型的一个版本。这里有系统完整训练的版本 (Example Models[7]).
通过上一节中的附加扩展,OpenNMT-py复制在EN-DE WMT上达到26.9。在这里,我已将这些参数加载到我们的重新实现中。
!wget https://s3.amazonaws.com/opennmt-models/en-de-model.pt
model, SRC, TGT = torch.load("en-de-model.pt")
model.eval()
sent = "▁The ▁log ▁file ▁can ▁be ▁sent ▁secret ly ▁with ▁email ▁or ▁FTP ▁to ▁a ▁specified ▁receiver".split()
src = torch.LongTensor([[SRC.stoi[w] for w in sent]])
src = Variable(src)
src_mask = (src != SRC.stoi["<blank>"]).unsqueeze(-2)
out = greedy_decode(model, src, src_mask,
max_len=60, start_symbol=TGT.stoi["<s>"])
print("Translation:", end="\t")
trans = "<s> "
for i in range(1, out.size(1)):
sym = TGT.itos[out[0, i]]
if sym == "</s>": break
trans += sym + " "
print(trans)
Translation: <s> ▁Die ▁Protokoll datei ▁kann ▁ heimlich ▁per ▁E - Mail ▁oder ▁FTP ▁an ▁einen ▁bestimmte n ▁Empfänger ▁gesendet ▁werden .
即使使用贪婪的解码器,翻译看起来也不错。我们可以进一步想象它,看看每一层注意力发生了什么。
tgt_sent = trans.split()
def draw(data, x, y, ax):
seaborn.heatmap(data,
xticklabels=x, square=True, yticklabels=y, vmin=0.0, vmax=1.0,
cbar=False, ax=ax)
for layer in range(1, 6, 2):
fig, axs = plt.subplots(1,4, figsize=(20, 10))
print("Encoder Layer", layer+1)
for h in range(4):
draw(model.encoder.layers[layer].self_attn.attn[0, h].data,
sent, sent if h ==0 else [], ax=axs[h])
plt.show()
for layer in range(1, 6, 2):
fig, axs = plt.subplots(1,4, figsize=(20, 10))
print("Decoder Self Layer", layer+1)
for h in range(4):
draw(model.decoder.layers[layer].self_attn.attn[0, h].data[:len(tgt_sent), :len(tgt_sent)],
tgt_sent, tgt_sent if h ==0 else [], ax=axs[h])
plt.show()
print("Decoder Src Layer", layer+1)
fig, axs = plt.subplots(1,4, figsize=(20, 10))
for h in range(4):
draw(model.decoder.layers[layer].self_attn.attn[0, h].data[:len(tgt_sent), :len(sent)],
sent, tgt_sent if h ==0 else [], ax=axs[h])
plt.show()
Encoder Layer 2
Encoder Layer 4
Encoder Layer 6
Decoder Self Layer 2
Decoder Src Layer 2
Decoder Self Layer 4
Decoder Src Layer 4
Decoder Self Layer 6
Decoder Src Layer 6
希望这段代码对未来的研究很有用。如果您有任何问题,请与我们联系。如果您发现此代码有用,请查看我们的其他OpenNMT工具。
@inproceedings{opennmt,
author = {Guillaume Klein and
Yoon Kim and
Yuntian Deng and
Jean Senellart and
Alexander M. Rush},
title = {OpenNMT: Open-Source Toolkit for Neural Machine Translation},
booktitle = {Proc. ACL},
year = {2017},
url = {https://doi.org/10.18653/v1/P17-4012},
doi = {10.18653/v1/P17-4012}
}
Cheers,srush
[1] https://arxiv.org/abs/1412.6980
[2] https://arxiv.org/abs/1512.00567
[3] https://github.com/opennmt/opennmt-py
[4] https://github.com/rsennrich/subword-nmt
[5] https://arxiv.org/abs/1608.05859
[6https://github.com/OpenNMT/OpenNM-py/blob/master/onmt/translate/Beam.py
[7] http://opennmt.net/Models-py/
本期责任编辑:张伟男
本期编辑:刘元兴
推荐阅读
征稿启示| 200元稿费+5000DBC(价值20个小时GPU算力)
完结撒花!李宏毅老师深度学习与人类语言处理课程视频及课件(附下载)
模型压缩实践系列之——bert-of-theseus,一个非常亲民的bert压缩方法
文本自动摘要任务的“不完全”心得总结番外篇——submodular函数优化
斯坦福大学NLP组Python深度学习自然语言处理工具Stanza试用
关于AINLP
AINLP 是一个有趣有AI的自然语言处理社区,专注于 AI、NLP、机器学习、深度学习、推荐算法等相关技术的分享,主题包括文本摘要、智能问答、聊天机器人、机器翻译、自动生成、知识图谱、预训练模型、推荐系统、计算广告、招聘信息、求职经验分享等,欢迎关注!加技术交流群请添加AINLPer(id:ainlper),备注工作/研究方向+加群目的。
阅读至此了,分享、点赞、在看三选一吧🙏