Class: Transformers::Vit::ViTSelfAttention

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

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ ViTSelfAttention

Returns a new instance of ViTSelfAttention.



102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
 
# File 'lib/transformers/models/vit/modeling_vit.rb', line 102

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

  @query = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: config.qkv_bias)
  @key = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: config.qkv_bias)
  @value = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: config.qkv_bias)

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

Instance Method Details

#forward(hidden_states, head_mask: nil, output_attentions: false) ⇒ 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
 
# File 'lib/transformers/models/vit/modeling_vit.rb', line 127

def forward(
  hidden_states, head_mask: nil, output_attentions: false
)
  mixed_query_layer = @query.(hidden_states)

  key_layer = transpose_for_scores(@key.(hidden_states))
  value_layer = transpose_for_scores(@value.(hidden_states))
  query_layer = transpose_for_scores(mixed_query_layer)

  # Take the dot product between "query" and "key" to get the raw attention scores.
  attention_scores = Torch.matmul(query_layer, key_layer.transpose(-1, -2))

  attention_scores = attention_scores / Math.sqrt(@attention_head_size)

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

  # This is actually dropping out entire tokens to attend to, which might
  # seem a bit unusual, but is taken from the original Transformer paper.
  attention_probs = @dropout.(attention_probs)

  # Mask heads if we want to
  if !head_mask.nil?
    attention_probs = attention_probs * head_mask
  end

  context_layer = Torch.matmul(attention_probs, value_layer)

  context_layer = context_layer.permute(0, 2, 1, 3).contiguous
  new_context_layer_shape = context_layer.size[...-2] + [@all_head_size]
  context_layer = context_layer.view(new_context_layer_shape)

  outputs = output_attentions ? [context_layer, attention_probs] : [context_layer]

  outputs
end

#transpose_for_scores(x) ⇒ Object 



121
122
123
124
125
 
# File 'lib/transformers/models/vit/modeling_vit.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