Class: TorchText::NN::MultiheadAttentionContainer

Inherits:
Torch::NN::Module
  • Object
show all
Defined in:
lib/torchtext/nn/multihead_attention_container.rb

Instance Method Summary collapse

Constructor Details

#initialize(nhead, in_proj_container, attention_layer, out_proj, batch_first: false) ⇒ MultiheadAttentionContainer 



4
5
6
7
8
9
10
11
 
# File 'lib/torchtext/nn/multihead_attention_container.rb', line 4

def initialize(nhead, in_proj_container, attention_layer, out_proj, batch_first: false)
  super()
  @nhead = nhead
  @in_proj_container = in_proj_container
  @attention_layer = attention_layer
  @out_proj = out_proj
  @batch_first = batch_first
end

Instance Method Details

#forward(query, key, value, attn_mask: nil, bias_k: nil, bias_v: nil) ⇒ Object 



13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
 
# File 'lib/torchtext/nn/multihead_attention_container.rb', line 13

def forward(query, key, value, attn_mask: nil, bias_k: nil, bias_v: nil)
  if @batch_first
    query, key, value = query.transpose(-3, -2), key.transpose(-3, -2), value.transpose(-3, -2)
  end

  tgt_len, src_len, bsz, embed_dim = query.size(-3), key.size(-3), query.size(-2), query.size(-1)
  q, k, v = @in_proj_container.call(query, key, value)
  unless q.size(-1) % @nhead == 0
    raise "query's embed_dim must be divisible by the number of heads"
  end
  head_dim = q.size(-1).div(@nhead)
  q = q.reshape(tgt_len, bsz * @nhead, head_dim)

  unless k.size(-1) % @nhead == 0
    raise "key's embed_dim must be divisible by the number of heads"
  end
  head_dim = k.size(-1).div(@nhead)
  k = k.reshape(src_len, bsz * @nhead, head_dim)

  unless v.size(-1) % @nhead == 0
    raise "value's embed_dim must be divisible by the number of heads"
  end
  head_dim = v.size(-1).div(@nhead)
  v = v.reshape(src_len, bsz * @nhead, head_dim)

  attn_output, attn_output_weights = @attention_layer.call(q, k, v, attn_mask: attn_mask, bias_k: bias_k, bias_v: bias_v)
  attn_output = attn_output.reshape(tgt_len, bsz, embed_dim)
  attn_output = @out_proj.call(attn_output)

  if @batch_first
    attn_output = attn_output.transpose(-3, -2)
  end

  [attn_output, attn_output_weights]
end