jupyterjazz commited on
Commit
ee8863c
1 Parent(s): 4b000ec

feat: matryoshka embeddings

Browse files

Signed-off-by: jupyterjazz <saba.sturua@jina.ai>

configuration_xlm_roberta.py CHANGED
@@ -31,6 +31,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
31
  use_flash_attn=True,
32
  torch_dtype=None,
33
  emb_pooler=None,
 
34
  **kwargs,
35
  ):
36
  super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
@@ -59,6 +60,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
59
  self.lora_main_params_trainable = lora_main_params_trainable
60
  self.use_flash_attn = use_flash_attn
61
  self.emb_pooler = emb_pooler
 
62
  if torch_dtype and hasattr(torch, torch_dtype) and type(getattr(torch, torch_dtype)) is torch.dtype:
63
  self.torch_dtype = getattr(torch, torch_dtype)
64
  else:
 
31
  use_flash_attn=True,
32
  torch_dtype=None,
33
  emb_pooler=None,
34
+ matryoshka_dimensions=None,
35
  **kwargs,
36
  ):
37
  super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
 
60
  self.lora_main_params_trainable = lora_main_params_trainable
61
  self.use_flash_attn = use_flash_attn
62
  self.emb_pooler = emb_pooler
63
+ self.matryoshka_dimensions = matryoshka_dimensions
64
  if torch_dtype and hasattr(torch, torch_dtype) and type(getattr(torch, torch_dtype)) is torch.dtype:
65
  self.torch_dtype = getattr(torch, torch_dtype)
66
  else:
modeling_xlm_roberta.py CHANGED
@@ -452,6 +452,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
452
  convert_to_tensor: bool = False,
453
  device: Optional[torch.device] = None,
454
  normalize_embeddings: bool = False,
 
455
  **tokenizer_kwargs,
456
  ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
457
  """
@@ -481,6 +482,8 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
481
  If set to true, returned vectors will have length 1. In that case, the
482
  faster dot-product (util.dot_score) instead of cosine similarity can
483
  be used.
 
 
484
  tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
485
  Keyword arguments for the tokenizer
486
  Returns:
@@ -575,6 +578,17 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
575
 
576
  all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
577
 
 
 
 
 
 
 
 
 
 
 
 
578
  if convert_to_tensor:
579
  all_embeddings = torch.stack(all_embeddings)
580
  elif convert_to_numpy:
 
452
  convert_to_tensor: bool = False,
453
  device: Optional[torch.device] = None,
454
  normalize_embeddings: bool = False,
455
+ truncate_dim: int = None,
456
  **tokenizer_kwargs,
457
  ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
458
  """
 
482
  If set to true, returned vectors will have length 1. In that case, the
483
  faster dot-product (util.dot_score) instead of cosine similarity can
484
  be used.
485
+ truncate_dim(`int`, *optional*, defaults to None):
486
+ The dimension to truncate sentence embeddings to. `None` does no truncation.
487
  tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
488
  Keyword arguments for the tokenizer
489
  Returns:
 
578
 
579
  all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
580
 
581
+ if truncate_dim:
582
+ if not self.config.matryoshka_dimension:
583
+ logger.warning(
584
+ 'Matryoshka embeddings are not supported, so dimension truncation will not be performed.'
585
+ )
586
+ elif truncate_dim in self.config.matryoshka_dimension:
587
+ all_embeddings = [tensor[:truncate_dim] for tensor in all_embeddings]
588
+ else:
589
+ raise ValueError(f'The provided `truncate_dim` value of {truncate_dim} is not supported. '
590
+ f'Supported dimensions are {self.config.matryoshka_dimension}.')
591
+
592
  if convert_to_tensor:
593
  all_embeddings = torch.stack(all_embeddings)
594
  elif convert_to_numpy: