Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Support embedding factorization in roberta #30528

Closed
wants to merge 2 commits into from
Closed

Support embedding factorization in roberta #30528

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cd761c4
Support embedding factorization in roberta
Apr 28, 2024
0d490c4
add factorized embedding to Bert
ming0308uk Apr 28, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/transformers/models/bert/configuration_bert.py
View file
Open in desktop
Expand Up @@ -116,6 +116,7 @@ def __init__(
position_embedding_type="absolute",
use_cache=True,
classifier_dropout=None,
embedding_size=None,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, **kwargs)
Expand All @@ -135,6 +136,7 @@ def __init__(
self.position_embedding_type = position_embedding_type
self.use_cache = use_cache
self.classifier_dropout = classifier_dropout
self.embedding_size = embedding_size


class BertOnnxConfig(OnnxConfig):
Expand Down
17 changes: 17 additions & 0 deletions src/transformers/models/bert/modeling_bert.py
View file
Open in desktop
Expand Up @@ -179,6 +179,9 @@ def __init__(self, config):
self.register_buffer(
"token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
)
self.embedding_size = config.embedding_size if config.embedding_size else config.hidden_size
if self.embedding_size != config.hidden_size:
self.embedding_transformation = nn.Linear(self.embedding_size, config.hidden_size)

def forward(
self,
Expand Down Expand Up @@ -211,6 +214,8 @@ def forward(

if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
if self.embedding_transformation:
inputs_embeds = self.embedding_transformation(inputs_embeds)
token_type_embeddings = self.token_type_embeddings(token_type_ids)

embeddings = inputs_embeds + token_type_embeddings
Expand Down Expand Up @@ -2012,3 +2017,15 @@ def forward(
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)

class NoNorm(nn.Module):
def __init__(self, feat_size, eps=None):
super().__init__()
self.bias = nn.Parameter(torch.zeros(feat_size))
self.weight = nn.Parameter(torch.ones(feat_size))

def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
return input_tensor * self.weight + self.bias


NORM2FN = {"layer_norm": nn.LayerNorm, "no_norm": NoNorm}
4 changes: 4 additions & 0 deletions src/transformers/models/roberta/configuration_roberta.py
View file
Open in desktop
Expand Up @@ -80,6 +80,8 @@ class RobertaConfig(PretrainedConfig):
relevant if `config.is_decoder=True`.
classifier_dropout (`float`, *optional*):
The dropout ratio for the classification head.
embedding_size (`int`, *optional*):
The dimension of the vocab embedding. If not set, it is the same as `hidden_size`.

Examples:

Expand Down Expand Up @@ -118,6 +120,7 @@ def __init__(
position_embedding_type="absolute",
use_cache=True,
classifier_dropout=None,
embedding_size=None,
**kwargs,
):
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
Expand All @@ -137,6 +140,7 @@ def __init__(
self.position_embedding_type = position_embedding_type
self.use_cache = use_cache
self.classifier_dropout = classifier_dropout
self.embedding_size = embedding_size


class RobertaOnnxConfig(OnnxConfig):
Expand Down
6 changes: 6 additions & 0 deletions src/transformers/models/roberta/modeling_roberta.py
View file
Open in desktop
Expand Up @@ -85,6 +85,8 @@ def __init__(self, config):
self.position_embeddings = nn.Embedding(
config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx
)
self.embedding_size = config.embedding_size if config.embedding_size else config.hidden_size
self.embedding_transformation = nn.Linear(config.embedding_size, config.hidden_size)

def forward(
self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
Expand Down Expand Up @@ -116,6 +118,10 @@ def forward(

if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)

if self.embedding_size != self.hidden_size:
inputs_embeds = self.embedding_transformation(inputs_embeds)

token_type_embeddings = self.token_type_embeddings(token_type_ids)

embeddings = inputs_embeds + token_type_embeddings
Expand Down