Class: TorchText::NN::ScaledDotProduct

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

Instance Method Summary collapse

Constructor Details

#initialize(dropout: 0.0, batch_first: false) ⇒ ScaledDotProduct

Returns a new instance of ScaledDotProduct.



4
5
6
7
8
 
# File 'lib/torchtext/nn/scaled_dot_product.rb', line 4

def initialize(dropout: 0.0, batch_first: false)
  super()
  @dropout = dropout
  @batch_first = batch_first
end

Instance Method Details

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



10
11
12
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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
 
# File 'lib/torchtext/nn/scaled_dot_product.rb', line 10

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

  if !bias_k.nil? && !bias_v.nil?
    unless key.size(-1) == bias_k.size(-1) && key.size(-2) == bias_k.size(-2) && bias_k.size(-3) == 1
      raise "Shape of bias_k is not supported"
    end
    unless value.size(-1) == bias_v.size(-1) && value.size(-2) == bias_v.size(-2) && bias_v.size(-3) == 1
      raise "Shape of bias_v is not supported"
    end
    key = Torch.cat([key, bias_k])
    value = Torch.cat([value, bias_v])
    if !attn_mask.nil?
      attn_mask = Torch::NN::Functional.pad(attn_mask, [0, 1])
    end
  end

  tgt_len, head_dim = query.size(-3), query.size(-1)
  unless query.size(-1) == key.size(-1) && key.size(-1) == value.size(-1)
    raise "The feature dim of query, key, value must be equal."
  end
  unless key.size() == value.size()
    raise "Shape of key, value must match"
  end
  src_len = key.size(-3)
  batch_heads = [query.size(-2), key.size(-2)].max

  # Scale query
  query, key, value = query.transpose(-2, -3), key.transpose(-2, -3), value.transpose(-2, -3)
  query = query * (head_dim.to_f ** -0.5)
  if !attn_mask.nil?
    if attn_mask.dim() != 3
      raise RuntimeError, "attn_mask must be a 3D tensor."
    end
    if (attn_mask.size(-1) != src_len) || (attn_mask.size(-2) != tgt_len) || (attn_mask.size(-3) != 1 && attn_mask.size(-3) != batch_heads)
      raise RuntimeError, "The size of the attn_mask is not correct."
    end
    if attn_mask.dtype != :bool
      raise RuntimeError, "Only bool tensor is supported for attn_mask"
    end
  end

  # Dot product of q, k
  attn_output_weights = Torch.matmul(query, key.transpose(-2, -1))
  if !attn_mask.nil?
    # TODO confirm last argument
    attn_output_weights.masked_fill!(attn_mask, -1e8, nil)
  end
  attn_output_weights = Torch::NN::Functional.softmax(attn_output_weights, dim: -1)
  attn_output_weights = Torch::NN::Functional.dropout(attn_output_weights, p: @dropout, training: @training)
  attn_output = Torch.matmul(attn_output_weights, value)

  if @batch_first
    [attn_output, attn_output_weights]
  else
    [attn_output.transpose(-3, -2), attn_output_weights]
  end
end