Class: Transformers::Mpnet::MPNetSelfAttention

Inherits:
Torch::NN::Module
  • Object
show all
Defined in:
lib/transformers/models/mpnet/modeling_mpnet.rb

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ MPNetSelfAttention

Returns a new instance of MPNetSelfAttention.



103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 103

def initialize(config)
  super()
  if config.hidden_size % config.num_attention_heads != 0 && !config.instance_variable_defined?(:@embedding_size)
    raise ArgumentError, "The hidden size (#{config.hidden_size}) is not a multiple of the number of attention heads (#{config.num_attention_heads})"
  end

  @num_attention_heads = config.num_attention_heads
  @attention_head_size = (config.hidden_size / config.num_attention_heads).to_i
  @all_head_size = @num_attention_heads * @attention_head_size

  @q = Torch::NN::Linear.new(config.hidden_size, @all_head_size)
  @k = Torch::NN::Linear.new(config.hidden_size, @all_head_size)
  @v = Torch::NN::Linear.new(config.hidden_size, @all_head_size)
  @o = Torch::NN::Linear.new(config.hidden_size, config.hidden_size)

  @dropout = Torch::NN::Dropout.new(p: config.attention_probs_dropout_prob)
end

Instance Method Details

#forward(hidden_states, attention_mask: nil, head_mask: nil, position_bias: nil, output_attentions: false, **kwargs) ⇒ Object 



127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 127

def forward(
  hidden_states,
  attention_mask: nil,
  head_mask: nil,
  position_bias: nil,
  output_attentions: false,
  **kwargs
)
  q = @q.(hidden_states)
  k = @k.(hidden_states)
  v = @v.(hidden_states)

  q = transpose_for_scores(q)
  k = transpose_for_scores(k)
  v = transpose_for_scores(v)

  # Take the dot product between "query" and "key" to get the raw attention scores.
  attention_scores = Torch.matmul(q, k.transpose(-1, -2))
  attention_scores = attention_scores / Math.sqrt(@attention_head_size)

  # Apply relative position embedding (precomputed in MPNetEncoder) if provided.
  if !position_bias.nil?
    attention_scores += position_bias
  end

  if !attention_mask.nil?
    attention_scores = attention_scores + attention_mask
  end

  # Normalize the attention scores to probabilities.
  attention_probs = Torch::NN::Functional.softmax(attention_scores, dim: -1)

  attention_probs = @dropout.(attention_probs)

  if !head_mask.nil?
    attention_probs = attention_probs * head_mask
  end

  c = Torch.matmul(attention_probs, v)

  c = c.permute(0, 2, 1, 3).contiguous
  new_c_shape = c.size[...-2] + [@all_head_size]
  c = c.view(*new_c_shape)

  o = @o.(c)

  outputs = output_attentions ? [o, attention_probs] : [o]
  outputs
end

#transpose_for_scores(x) ⇒ Object 



121
122
123
124
125
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 121

def transpose_for_scores(x)
  new_x_shape = x.size[...-1] + [@num_attention_heads, @attention_head_size]
  x = x.view(*new_x_shape)
  x.permute(0, 2, 1, 3)
end