Multi-head self-attention

Transformermodeller med PyTorch

James Chapman

Curriculum Manager, DataCamp

Multi-head attention i transformers

Transformerarkitekturen som visas i den akademiska artikeln Attention Is All You Need.

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Q: Anger vad varje token "letar efter" hos andra tokens
K: Representerar innehållet i varje token
V: Det faktiska innehåll som aggregeras eller viktas

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Q: Anger vad varje token "letar efter" hos andra tokens
K: Representerar innehållet i varje token
V: Det faktiska innehåll som aggregeras eller viktas

Attention Scores: Q-K-likhet → skalärprodukt

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Q: Anger vad varje token "letar efter" hos andra tokens
K: Representerar innehållet i varje token
V: Det faktiska innehåll som aggregeras eller viktas

Attention Scores: Q-K-likhet → skalärprodukt
Attention Weights: softmax-skalning

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Attention Scores: Q-K-likhet → skalärprodukt
Attention Weights: softmax-skalning

	Orange	is	my	favorite	fruit
Query:	Orange
Attention weights:	.21	.03	.05	.31	.40

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Attention Scores: Q-K-likhet → skalärprodukt
Attention Weights: softmax-skalning

	Orange	is	my	favorite	fruit
Query:	Orange
Attention weights:	.21	.03	.05	.31	.40

Self-attention-mekanismen

Anatomi för self-attention-mekanismen

Attention Scores: Q-K-likhet → skalärprodukt
Attention Weights: softmax-skalning

	Orange	is	my	favorite	fruit
Query:	Orange
Attention weights:	.21	.03	.05	.31	.40

Multi-head attention

Multi-headed self-attention

Multi-head attention

Multi-headed self-attention

Multi-head attention

Multi-headed self-attention

import torch.nn as nn
import torch.nn.functional as F

class MultiHeadAttention(nn.Module):

    def __init__(self, d_model, num_heads):
        super().__init__()

        assert d_model % num_heads == 0, "d_model must be divisible by num_heads."

        self.num_heads = num_heads
        self.d_model = d_model
        self.head_dim = d_model // num_heads

        self.query_linear = nn.Linear(d_model, d_model, bias=False)
        self.key_linear = nn.Linear(d_model, d_model, bias=False)
        self.value_linear = nn.Linear(d_model, d_model, bias=False)

        self.output_linear = nn.Linear(d_model, d_model)

num_heads: antal attention-huvuden, vart och ett hanterar inbäddningar av storleken head_dim
bias=False: ingen påverkan på prestanda men minskar komplexiteten (endast för indata)

    def split_heads(self, x, batch_size):

        seq_length = x.size(1)
        x = x.reshape(batch_size, seq_length, self.num_heads, self.head_dim)
        return x.permute(0, 2, 1, 3)


    def compute_attention(self, query, key, value, mask=None):

        scores = torch.matmul(query, key.transpose(-2, -1)) / (self.head_dim ** 0.5)

        if mask is not None:
            scores = scores.masked_fill(mask == 0, float('-inf'))

        attention_weights = F.softmax(scores, dim=-1)
        return torch.matmul(attention_weights, value)


    def combine_heads(self, x, batch_size):

        x = x.permute(0, 2, 1, 3).contiguous()
        return x.view(batch_size, -1, self.d_model)

compute_attention(): beräknar attention-vikter med F.softmax()
torch.matmul(attention_weights, value): viktad summa av värden

    def forward(self, query, key, value, mask=None):
        batch_size = query.size(0)

        query = self.split_heads(self.query_linear(query), batch_size)
        key = self.split_heads(self.key_linear(key), batch_size)
        value = self.split_heads(self.value_linear(value), batch_size)

        attention_weights = self.compute_attention(query, key, value, mask)


        output = self.combine_heads(attention_weights, batch_size)

        return self.output_linear(output)

self.output_linear(): konkatenerar och projicerar huvudutdata

Nu kör vi en övning!

Transformermodeller med PyTorch