Class: Transformers::XlmRoberta::XLMRobertaEmbeddings

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

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ XLMRobertaEmbeddings

Returns a new instance of XLMRobertaEmbeddings.



19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
 
# File 'lib/transformers/models/xlm_roberta/modeling_xlm_roberta.rb', line 19

def initialize(config)
  super()
  @word_embeddings = Torch::NN::Embedding.new(config.vocab_size, config.hidden_size, padding_idx: config.pad_token_id)
  @position_embeddings = Torch::NN::Embedding.new(config.max_position_embeddings, config.hidden_size)
  @token_type_embeddings = Torch::NN::Embedding.new(config.type_vocab_size, config.hidden_size)

  # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
  # any TensorFlow checkpoint file
  @LayerNorm = Torch::NN::LayerNorm.new(config.hidden_size, eps: config.layer_norm_eps)
  @dropout = Torch::NN::Dropout.new(p: config.hidden_dropout_prob)
  # position_ids (1, len position emb) is contiguous in memory and exported when serialized
  @position_embedding_type = config.getattr("position_embedding_type", "absolute")
  register_buffer("position_ids", Torch.arange(config.max_position_embeddings).expand([1, -1]), persistent: false)
  register_buffer("token_type_ids", Torch.zeros(@position_ids.size, dtype: Torch.long), persistent: false)

  @padding_idx = config.pad_token_id
  @position_embeddings = Torch::NN::Embedding.new(config.max_position_embeddings, config.hidden_size, padding_idx: @padding_idx)
end

Instance Method Details

#create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length: 0) ⇒ Object 



92
93
94
95
96
97
 
# File 'lib/transformers/models/xlm_roberta/modeling_xlm_roberta.rb', line 92

def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length: 0)
  # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
  mask = input_ids.ne(padding_idx).int
  incremental_indices = (Torch.cumsum(mask, dim: 1).type_as(mask) + past_key_values_length) * mask
  incremental_indices.long + padding_idx
end

#create_position_ids_from_inputs_embeds(inputs_embeds) ⇒ Object 



84
85
86
87
88
89
90
 
# File 'lib/transformers/models/xlm_roberta/modeling_xlm_roberta.rb', line 84

def create_position_ids_from_inputs_embeds(inputs_embeds)
  input_shape = inputs_embeds.size[...-1]
  sequence_length = input_shape[1]

  position_ids = Torch.arange(@padding_idx + 1, sequence_length + @padding_idx + 1, dtype: Torch.long, device: inputs_embeds.device)
  position_ids.unsqueeze(0).expand(input_shape)
end

#forward(input_ids: nil, token_type_ids: nil, position_ids: nil, inputs_embeds: nil, past_key_values_length: 0) ⇒ Object 



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
70
71
72
73
74
75
76
77
78
79
80
81
82
 
# File 'lib/transformers/models/xlm_roberta/modeling_xlm_roberta.rb', line 38

def forward(input_ids: nil, token_type_ids: nil, position_ids: nil, inputs_embeds: nil, past_key_values_length: 0)
  if position_ids.nil?
    if !input_ids.nil?
      # Create the position ids from the input token ids. Any padded tokens remain padded.
      position_ids = create_position_ids_from_input_ids(input_ids, @padding_idx, past_key_values_length:)
    else
      position_ids = create_position_ids_from_inputs_embeds(inputs_embeds)
    end
  end

  if !input_ids.nil?
    input_shape = input_ids.size
  else
    input_shape = inputs_embeds.size[...-1]
  end

  seq_length = input_shape[1]

  # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
  # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
  # issue #5664
  if token_type_ids.nil?
    if respond_to?(:token_type_ids)
      buffered_token_type_ids = token_type_ids[0.., ...seq_length]
      buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
      token_type_ids = buffered_token_type_ids_expanded
    else
      token_type_ids = Torch.zeros(input_shape, dtype: Torch.long, device: @position_ids.device)
    end
  end

  if inputs_embeds.nil?
    inputs_embeds = @word_embeddings.(input_ids)
  end
  token_type_embeddings = @token_type_embeddings.(token_type_ids)

  embeddings = inputs_embeds + token_type_embeddings
  if @position_embedding_type == "absolute"
    position_embeddings = @position_embeddings.(position_ids)
    embeddings += position_embeddings
  end
  embeddings = @LayerNorm.(embeddings)
  embeddings = @dropout.(embeddings)
  embeddings
end