.. _sec_natural-language-inference-attention:
Inferência de Linguagem Natural: Usando a Atenção
=================================================
Introduzimos a tarefa de inferência em linguagem natural e o conjunto de
dados SNLI em :numref:`sec_natural-language-inference-and-dataset`. Em
vista de muitos modelos baseados em arquiteturas complexas e profundas,
Parikh et al. proposto para abordar a inferência de linguagem natural
com mecanismos de atenção e chamou-o de “modelo de atenção decomposto”
:cite:`Parikh.Tackstrom.Das.ea.2016`. Isso resulta em um modelo sem
camadas recorrentes ou convolucionais, alcançando o melhor resultado no
momento no conjunto de dados SNLI com muito menos parâmetros. Nesta
seção, iremos descrever e implementar este método baseado em atenção
(com MLPs) para inferência de linguagem natural, conforme descrito em
:numref:`fig_nlp-map-nli -ention`.
.. _fig_nlp-map-nli-attention:
.. figure:: ../img/nlp-map-nli-attention.svg
Esta seção alimenta o GloVe pré-treinado para uma arquitetura baseada
em atenção e MLPs para inferência de linguagem natural.
O Modelo
--------
Mais simples do que preservar a ordem das palavras em premissas e
hipóteses, podemos apenas alinhar as palavras em uma sequência de texto
com todas as palavras na outra e vice-versa, em seguida, compare e
agregue essas informações para prever as relações lógicas entre
premissas e hipóteses. Semelhante ao alinhamento de palavras entre as
frases fonte e alvo na tradução automática, o alinhamento de palavras
entre premissas e hipóteses pode ser perfeitamente realizado por
mecanismos de atenção.
.. _fig_nli_attention:
.. figure:: ../img/nli-attention.svg
Inferência de linguagem natural usando mecanismos de atenção.
:numref:`fig_nli_attention` descreve o método de inferência de
linguagem natural usando mecanismos de atenção. Em um nível superior,
consiste em três etapas treinadas em conjunto: alinhar, comparar e
agregar. Iremos ilustrá-los passo a passo a seguir.
.. raw:: html
.. raw:: html
.. code:: python
from mxnet import gluon, init, np, npx
from mxnet.gluon import nn
from d2l import mxnet as d2l
npx.set_np()
.. raw:: html
.. raw:: html
.. code:: python
import torch
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
def mlp(num_hiddens, flatten):
net = nn.Sequential()
net.add(nn.Dropout(0.2))
net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))
net.add(nn.Dropout(0.2))
net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))
return net
.. raw:: html
.. raw:: html
.. code:: python
def mlp(num_inputs, num_hiddens, flatten):
net = []
net.append(nn.Dropout(0.2))
net.append(nn.Linear(num_inputs, num_hiddens))
net.append(nn.ReLU())
if flatten:
net.append(nn.Flatten(start_dim=1))
net.append(nn.Dropout(0.2))
net.append(nn.Linear(num_hiddens, num_hiddens))
net.append(nn.ReLU())
if flatten:
net.append(nn.Flatten(start_dim=1))
return nn.Sequential(*net)
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class Attend(nn.Block):
def __init__(self, num_hiddens, **kwargs):
super(Attend, self).__init__(**kwargs)
self.f = mlp(num_hiddens=num_hiddens, flatten=False)
def forward(self, A, B):
# Shape of `A`/`B`: (b`atch_size`, no. of words in sequence A/B,
# `embed_size`)
# Shape of `f_A`/`f_B`: (`batch_size`, no. of words in sequence A/B,
# `num_hiddens`)
f_A = self.f(A)
f_B = self.f(B)
# Shape of `e`: (`batch_size`, no. of words in sequence A,
# no. of words in sequence B)
e = npx.batch_dot(f_A, f_B, transpose_b=True)
# Shape of `beta`: (`batch_size`, no. of words in sequence A,
# `embed_size`), where sequence B is softly aligned with each word
# (axis 1 of `beta`) in sequence A
beta = npx.batch_dot(npx.softmax(e), B)
# Shape of `alpha`: (`batch_size`, no. of words in sequence B,
# `embed_size`), where sequence A is softly aligned with each word
# (axis 1 of `alpha`) in sequence B
alpha = npx.batch_dot(npx.softmax(e.transpose(0, 2, 1)), A)
return beta, alpha
.. raw:: html
.. raw:: html
.. code:: python
class Attend(nn.Module):
def __init__(self, num_inputs, num_hiddens, **kwargs):
super(Attend, self).__init__(**kwargs)
self.f = mlp(num_inputs, num_hiddens, flatten=False)
def forward(self, A, B):
# Shape of `A`/`B`: (`batch_size`, no. of words in sequence A/B,
# `embed_size`)
# Shape of `f_A`/`f_B`: (`batch_size`, no. of words in sequence A/B,
# `num_hiddens`)
f_A = self.f(A)
f_B = self.f(B)
# Shape of `e`: (`batch_size`, no. of words in sequence A,
# no. of words in sequence B)
e = torch.bmm(f_A, f_B.permute(0, 2, 1))
# Shape of `beta`: (`batch_size`, no. of words in sequence A,
# `embed_size`), where sequence B is softly aligned with each word
# (axis 1 of `beta`) in sequence A
beta = torch.bmm(F.softmax(e, dim=-1), B)
# Shape of `alpha`: (`batch_size`, no. of words in sequence B,
# `embed_size`), where sequence A is softly aligned with each word
# (axis 1 of `alpha`) in sequence B
alpha = torch.bmm(F.softmax(e.permute(0, 2, 1), dim=-1), A)
return beta, alpha
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class Compare(nn.Block):
def __init__(self, num_hiddens, **kwargs):
super(Compare, self).__init__(**kwargs)
self.g = mlp(num_hiddens=num_hiddens, flatten=False)
def forward(self, A, B, beta, alpha):
V_A = self.g(np.concatenate([A, beta], axis=2))
V_B = self.g(np.concatenate([B, alpha], axis=2))
return V_A, V_B
.. raw:: html
.. raw:: html
.. code:: python
class Compare(nn.Module):
def __init__(self, num_inputs, num_hiddens, **kwargs):
super(Compare, self).__init__(**kwargs)
self.g = mlp(num_inputs, num_hiddens, flatten=False)
def forward(self, A, B, beta, alpha):
V_A = self.g(torch.cat([A, beta], dim=2))
V_B = self.g(torch.cat([B, alpha], dim=2))
return V_A, V_B
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class Aggregate(nn.Block):
def __init__(self, num_hiddens, num_outputs, **kwargs):
super(Aggregate, self).__init__(**kwargs)
self.h = mlp(num_hiddens=num_hiddens, flatten=True)
self.h.add(nn.Dense(num_outputs))
def forward(self, V_A, V_B):
# Sum up both sets of comparison vectors
V_A = V_A.sum(axis=1)
V_B = V_B.sum(axis=1)
# Feed the concatenation of both summarization results into an MLP
Y_hat = self.h(np.concatenate([V_A, V_B], axis=1))
return Y_hat
.. raw:: html
.. raw:: html
.. code:: python
class Aggregate(nn.Module):
def __init__(self, num_inputs, num_hiddens, num_outputs, **kwargs):
super(Aggregate, self).__init__(**kwargs)
self.h = mlp(num_inputs, num_hiddens, flatten=True)
self.linear = nn.Linear(num_hiddens, num_outputs)
def forward(self, V_A, V_B):
# Sum up both sets of comparison vectors
V_A = V_A.sum(dim=1)
V_B = V_B.sum(dim=1)
# Feed the concatenation of both summarization results into an MLP
Y_hat = self.linear(self.h(torch.cat([V_A, V_B], dim=1)))
return Y_hat
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class DecomposableAttention(nn.Block):
def __init__(self, vocab, embed_size, num_hiddens, **kwargs):
super(DecomposableAttention, self).__init__(**kwargs)
self.embedding = nn.Embedding(len(vocab), embed_size)
self.attend = Attend(num_hiddens)
self.compare = Compare(num_hiddens)
# There are 3 possible outputs: entailment, contradiction, and neutral
self.aggregate = Aggregate(num_hiddens, 3)
def forward(self, X):
premises, hypotheses = X
A = self.embedding(premises)
B = self.embedding(hypotheses)
beta, alpha = self.attend(A, B)
V_A, V_B = self.compare(A, B, beta, alpha)
Y_hat = self.aggregate(V_A, V_B)
return Y_hat
.. raw:: html
.. raw:: html
.. code:: python
class DecomposableAttention(nn.Module):
def __init__(self, vocab, embed_size, num_hiddens, num_inputs_attend=100,
num_inputs_compare=200, num_inputs_agg=400, **kwargs):
super(DecomposableAttention, self).__init__(**kwargs)
self.embedding = nn.Embedding(len(vocab), embed_size)
self.attend = Attend(num_inputs_attend, num_hiddens)
self.compare = Compare(num_inputs_compare, num_hiddens)
# There are 3 possible outputs: entailment, contradiction, and neutral
self.aggregate = Aggregate(num_inputs_agg, num_hiddens, num_outputs=3)
def forward(self, X):
premises, hypotheses = X
A = self.embedding(premises)
B = self.embedding(hypotheses)
beta, alpha = self.attend(A, B)
V_A, V_B = self.compare(A, B, beta, alpha)
Y_hat = self.aggregate(V_A, V_B)
return Y_hat
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
batch_size, num_steps = 256, 50
train_iter, test_iter, vocab = d2l.load_data_snli(batch_size, num_steps)
.. parsed-literal::
:class: output
read 549367 examples
read 9824 examples
.. raw:: html
.. raw:: html
.. code:: python
batch_size, num_steps = 256, 50
train_iter, test_iter, vocab = d2l.load_data_snli(batch_size, num_steps)
.. parsed-literal::
:class: output
read 549367 examples
read 9824 examples
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
embed_size, num_hiddens, devices = 100, 200, d2l.try_all_gpus()
net = DecomposableAttention(vocab, embed_size, num_hiddens)
net.initialize(init.Xavier(), ctx=devices)
glove_embedding = d2l.TokenEmbedding('glove.6b.100d')
embeds = glove_embedding[vocab.idx_to_token]
net.embedding.weight.set_data(embeds)
.. raw:: html
.. raw:: html
.. code:: python
embed_size, num_hiddens, devices = 100, 200, d2l.try_all_gpus()
net = DecomposableAttention(vocab, embed_size, num_hiddens)
glove_embedding = d2l.TokenEmbedding('glove.6b.100d')
embeds = glove_embedding[vocab.idx_to_token]
net.embedding.weight.data.copy_(embeds);
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
#@save
def split_batch_multi_inputs(X, y, devices):
"""Split multi-input `X` and `y` into multiple devices."""
X = list(zip(*[gluon.utils.split_and_load(
feature, devices, even_split=False) for feature in X]))
return (X, gluon.utils.split_and_load(y, devices, even_split=False))
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
lr, num_epochs = 0.001, 4
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': lr})
loss = gluon.loss.SoftmaxCrossEntropyLoss()
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices,
split_batch_multi_inputs)
.. parsed-literal::
:class: output
loss 0.514, train acc 0.797, test acc 0.815
7637.6 examples/sec on [gpu(0), gpu(1)]
.. figure:: output_natural-language-inference-attention_b907c4_81_1.svg
.. raw:: html
.. raw:: html
.. code:: python
lr, num_epochs = 0.001, 4
trainer = torch.optim.Adam(net.parameters(), lr=lr)
loss = nn.CrossEntropyLoss(reduction="none")
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices)
.. parsed-literal::
:class: output
loss 0.496, train acc 0.804, test acc 0.827
18580.5 examples/sec on [device(type='cuda', index=0), device(type='cuda', index=1)]
.. figure:: output_natural-language-inference-attention_b907c4_84_1.svg
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
#@save
def predict_snli(net, vocab, premise, hypothesis):
premise = np.array(vocab[premise], ctx=d2l.try_gpu())
hypothesis = np.array(vocab[hypothesis], ctx=d2l.try_gpu())
label = np.argmax(net([premise.reshape((1, -1)),
hypothesis.reshape((1, -1))]), axis=1)
return 'entailment' if label == 0 else 'contradiction' if label == 1 \
else 'neutral'
.. raw:: html
.. raw:: html
.. code:: python
#@save
def predict_snli(net, vocab, premise, hypothesis):
net.eval()
premise = torch.tensor(vocab[premise], device=d2l.try_gpu())
hypothesis = torch.tensor(vocab[hypothesis], device=d2l.try_gpu())
label = torch.argmax(net([premise.reshape((1, -1)),
hypothesis.reshape((1, -1))]), dim=1)
return 'entailment' if label == 0 else 'contradiction' if label == 1 \
else 'neutral'
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
predict_snli(net, vocab, ['he', 'is', 'good', '.'], ['he', 'is', 'bad', '.'])
.. parsed-literal::
:class: output
'contradiction'
.. raw:: html
.. raw:: html
.. code:: python
predict_snli(net, vocab, ['he', 'is', 'good', '.'], ['he', 'is', 'bad', '.'])
.. parsed-literal::
:class: output
'contradiction'
.. raw:: html
.. raw:: html
.. raw:: html
`Discussõess `__
.. raw:: html
.. raw:: html
`Discussõess `__
.. raw:: html
.. raw:: html