Class: Transformers::DebertaV2::DisentangledSelfAttention

Inherits:
Torch::NN::Module
  • Object
show all
Defined in:
lib/transformers/models/deberta_v2/modeling_deberta_v2.rb

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ DisentangledSelfAttention

Returns a new instance of DisentangledSelfAttention.



462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
 
# File 'lib/transformers/models/deberta_v2/modeling_deberta_v2.rb', line 462

def initialize(config)
  super()
  if config.hidden_size % config.num_attention_heads != 0
    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
  @attention_head_size = config.getattr("attention_head_size", _attention_head_size)
  @all_head_size = @num_attention_heads * @attention_head_size
  @query_proj = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: true)
  @key_proj = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: true)
  @value_proj = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: true)

  @share_att_key = config.getattr("share_att_key", false)
  @pos_att_type = !config.pos_att_type.nil? ? config.pos_att_type : []
  @relative_attention = config.getattr("relative_attention", false)

  if @relative_attention
    @position_buckets = config.getattr("position_buckets", -1)
    @max_relative_positions = config.getattr("max_relative_positions", -1)
    if @max_relative_positions < 1
      @max_relative_positions = config.max_position_embeddings
    end
    @pos_ebd_size = @max_relative_positions
    if @position_buckets > 0
      @pos_ebd_size = @position_buckets
    end

    @pos_dropout = StableDropout.new(config.hidden_dropout_prob)

    if !@share_att_key
      if @pos_att_type.include?("c2p")
        @pos_key_proj = Torch::NN::Linear.new(config.hidden_size, @all_head_size, bias: true)
      end
      if @pos_att_type.include?("p2c")
        @pos_query_proj = Torch::NN::Linear.new(config.hidden_size, @all_head_size)
      end
    end
  end

  @dropout = StableDropout.new(config.attention_probs_dropout_prob)
end

Instance Method Details

#disentangled_attention_bias(query_layer, key_layer, relative_pos, rel_embeddings, scale_factor) ⇒ Object 



562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
 
# File 'lib/transformers/models/deberta_v2/modeling_deberta_v2.rb', line 562

def disentangled_attention_bias(query_layer, key_layer, relative_pos, rel_embeddings, scale_factor)
  if relative_pos.nil?
    q = query_layer.size(-2)
    relative_pos = DebertaV2.build_relative_position(q, key_layer.size(-2), bucket_size: @position_buckets, max_position: @max_relative_positions, device: query_layer.device)
  end
  if relative_pos.dim == 2
    relative_pos = relative_pos.unsqueeze(0).unsqueeze(0)
  elsif relative_pos.dim == 3
    relative_pos = relative_pos.unsqueeze(1)
  elsif relative_pos.dim != 4
    raise ArgumentError, "Relative position ids must be of dim 2 or 3 or 4. #{relative_pos.dim}"
  end

  att_span = @pos_ebd_size
  relative_pos = relative_pos.long.to(query_layer.device)

  rel_embeddings = rel_embeddings[0...att_span * 2, 0..].unsqueeze(0)
  if @share_att_key
    pos_query_layer = transpose_for_scores(@query_proj.(rel_embeddings), @num_attention_heads).repeat(query_layer.size(0) / @num_attention_heads, 1, 1)
    pos_key_layer = transpose_for_scores(@key_proj.(rel_embeddings), @num_attention_heads).repeat(query_layer.size(0) / @num_attention_heads, 1, 1)
  elsif @pos_att_type.include?("c2p")
    pos_key_layer = transpose_for_scores(@pos_key_proj.(rel_embeddings), @num_attention_heads).repeat(query_layer.size(0) / @num_attention_heads, 1, 1)
  end

  score = 0
  # content->position
  if @pos_att_type.include?("c2p")
    scale = Torch.sqrt(Torch.tensor(pos_key_layer.size(-1), dtype: Torch.float) * scale_factor)
    c2p_att = Torch.bmm(query_layer, pos_key_layer.transpose(-1, -2))
    c2p_pos = Torch.clamp(relative_pos + att_span, 0, (att_span * 2) - 1)
    c2p_att = Torch.gather(c2p_att, dim: -1, index: c2p_pos.squeeze(0).expand([query_layer.size(0), query_layer.size(1), relative_pos.size(-1)]))
    score += c2p_att / scale.to(dtype: c2p_att.dtype)
  end

  # position->content
  if @pos_att_type.include?("p2c")
    scale = Torch.sqrt(Torch.tensor(pos_query_layer.size(-1), dtype: Torch.float) * scale_factor)
    if key_layer.size(-2) != query_layer.size(-2)
      r_pos = DebertaV2.build_relative_position(key_layer.size(-2), key_layer.size(-2), bucket_size: @position_buckets, max_position: @max_relative_positions, device: query_layer.device)
      r_pos = r_pos.unsqueeze(0)
    else
      r_pos = relative_pos
    end

    p2c_pos = Torch.clamp(-r_pos + att_span, 0, (att_span * 2) - 1)
    p2c_att = Torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
    p2c_att = Torch.gather(p2c_att, dim: -1, index: p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)])).transpose(-1, -2)
    score += p2c_att / scale.to(dtype: p2c_att.dtype)
  end

  score
end

#forward(hidden_states, attention_mask, output_attentions: false, query_states: nil, relative_pos: nil, rel_embeddings: nil) ⇒ Object 



511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
 
# File 'lib/transformers/models/deberta_v2/modeling_deberta_v2.rb', line 511

def forward(
  hidden_states,
  attention_mask,
  output_attentions: false,
  query_states: nil,
  relative_pos: nil,
  rel_embeddings: nil
)
  if query_states.nil?
    query_states = hidden_states
  end
  query_layer = transpose_for_scores(@query_proj.(query_states), @num_attention_heads)
  key_layer = transpose_for_scores(@key_proj.(hidden_states), @num_attention_heads)
  value_layer = transpose_for_scores(@value_proj.(hidden_states), @num_attention_heads)

  rel_att = nil
  # Take the dot product between "query" and "key" to get the raw attention scores.
  scale_factor = 1
  if @pos_att_type.include?("c2p")
    scale_factor += 1
  end
  if @pos_att_type.include?("p2c")
    scale_factor += 1
  end
  scale = Torch.sqrt(Torch.tensor(query_layer.size(-1), dtype: Torch.float) * scale_factor)
  attention_scores = Torch.bmm(query_layer, key_layer.transpose(-1, -2) / scale.to(dtype: query_layer.dtype))
  if @relative_attention
    rel_embeddings = @pos_dropout.(rel_embeddings)
    rel_att = disentangled_attention_bias(query_layer, key_layer, relative_pos, rel_embeddings, scale_factor)
  end

  if !rel_att.nil?
    attention_scores = attention_scores + rel_att
  end
  attention_scores = attention_scores
  attention_scores = attention_scores.view(-1, @num_attention_heads, attention_scores.size(-2), attention_scores.size(-1))

  # bsz x height x length x dimension
  attention_probs = XSoftmax.apply(attention_scores, attention_mask, -1)
  attention_probs = @dropout.(attention_probs)
  context_layer = Torch.bmm(attention_probs.view(-1, attention_probs.size(-2), attention_probs.size(-1)), value_layer)
  context_layer = context_layer.view(-1, @num_attention_heads, context_layer.size(-2), context_layer.size(-1)).permute(0, 2, 1, 3).contiguous
  new_context_layer_shape = context_layer.size[...-2] + [-1]
  context_layer = context_layer.view(new_context_layer_shape)
  if output_attentions
    [context_layer, attention_probs]
  else
    context_layer
  end
end

#transpose_for_scores(x, attention_heads) ⇒ Object 



505
506
507
508
509
 
# File 'lib/transformers/models/deberta_v2/modeling_deberta_v2.rb', line 505

def transpose_for_scores(x, attention_heads)
  new_x_shape = x.size[...-1] + [attention_heads, -1]
  x = x.view(new_x_shape)
  x.permute(0, 2, 1, 3).contiguous.view(-1, x.size(1), x.size(-1))
end