Class: Transformers::Mpnet::MPNetEncoder

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

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ MPNetEncoder

Returns a new instance of MPNetEncoder.



281
282
283
284
285
286
287
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 281

def initialize(config)
  super()
  @config = config
  @n_heads = config.num_attention_heads
  @layer = Torch::NN::ModuleList.new(config.num_hidden_layers.times.map { |_| MPNetLayer.new(config) })
  @relative_attention_bias = Torch::NN::Embedding.new(config.relative_attention_num_buckets, @n_heads)
end

Class Method Details

.relative_position_bucket(relative_position, num_buckets: 32, max_distance: 128) ⇒ Object 



345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 345

def self.relative_position_bucket(relative_position, num_buckets: 32, max_distance: 128)
  ret = 0
  n = -relative_position

  num_buckets /= 2
  ret += n.lt(0).to(Torch.long) * num_buckets
  n = Torch.abs(n)

  max_exact = num_buckets / 2
  is_small = n.lt(max_exact)

  val_if_large = max_exact + (
    Torch.log(n.float / max_exact) / Math.log(max_distance / max_exact) * (num_buckets - max_exact)
  ).to(Torch.long)

  val_if_large = Torch.min(val_if_large, Torch.full_like(val_if_large, num_buckets - 1))
  ret += Torch.where(is_small, n, val_if_large)
  ret
end

Instance Method Details

#compute_position_bias(x, position_ids: nil, num_buckets: 32) ⇒ Object 



325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 325

def compute_position_bias(x, position_ids: nil, num_buckets: 32)
  bsz, qlen, klen = [x.size(0), x.size(1), x.size(1)]
  if !position_ids.nil?
    context_position = position_ids[0.., 0.., nil]
    memory_position = position_ids[0.., nil, 0..]
  else
    context_position = Torch.arange(qlen, dtype: Torch.long)[0.., nil]
    memory_position = Torch.arange(klen, dtype: Torch.long)[nil, 0..]
  end

  relative_position = memory_position - context_position

  rp_bucket = self.class.relative_position_bucket(relative_position, num_buckets: num_buckets)
  rp_bucket = rp_bucket.to(x.device)
  values = @relative_attention_bias.(rp_bucket)
  values = values.permute([2, 0, 1]).unsqueeze(0)
  values = values.expand([bsz, -1, qlen, klen]).contiguous
  values
end

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



289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
 
# File 'lib/transformers/models/mpnet/modeling_mpnet.rb', line 289

def forward(
  hidden_states,
  attention_mask: nil,
  head_mask: nil,
  output_attentions: false,
  output_hidden_states: false,
  return_dict: false,
  **kwargs
)
  position_bias = compute_position_bias(hidden_states)
  all_hidden_states = output_hidden_states ? [] : nil
  all_attentions = output_attentions ? [] : nil
  @layer.each_with_index do |layer_module, i|
    if output_hidden_states
      all_hidden_states = all_hidden_states + [hidden_states]
    end

    layer_outputs = layer_module.(hidden_states, attention_mask: attention_mask, head_mask: head_mask[i], position_bias: position_bias, output_attentions: output_attentions, **kwargs)
    hidden_states = layer_outputs[0]

    if output_attentions
      all_attentions = all_attentions + [layer_outputs[1]]
    end
  end

  # Add last layer
  if output_hidden_states
    all_hidden_states = all_hidden_states + [hidden_states]
  end

  if !return_dict
    return Array([hidden_states, all_hidden_states, all_attentions].select { |v| !v.nil? })
  end
  BaseModelOutput.new(last_hidden_state: hidden_states, hidden_states: all_hidden_states, attentions: all_attentions)
end