michael-guenther
commited on
Commit
•
2b6bc3f
1
Parent(s):
54b019f
output-hidden-states (#56)
Browse files- add option to output hidden states (681845da60fc92b982a091b193ae1fac09bc61f8)
- modeling_xlm_roberta.py +43 -10
modeling_xlm_roberta.py
CHANGED
@@ -22,13 +22,13 @@ import torch.nn.functional as F
|
|
22 |
import torch.utils.checkpoint
|
23 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
24 |
from transformers import AutoTokenizer, PretrainedConfig
|
25 |
-
from transformers.modeling_outputs import
|
26 |
-
SequenceClassifierOutput)
|
27 |
from transformers.modeling_utils import PreTrainedModel
|
28 |
from transformers.models.bert.modeling_bert import (
|
29 |
-
BaseModelOutputWithPoolingAndCrossAttentions,
|
30 |
-
|
31 |
-
|
|
|
32 |
|
33 |
from .rotary import RotaryEmbedding
|
34 |
from .block import Block
|
@@ -195,17 +195,30 @@ class XLMRobertaEncoder(nn.Module):
|
|
195 |
self._grad_checkpointing = value
|
196 |
|
197 |
def forward(
|
198 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
199 |
):
|
200 |
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
201 |
This means that we only compute the last layer output for these tokens.
|
202 |
subset_mask: (batch, seqlen), dtype=torch.bool
|
203 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
204 |
if key_padding_mask is None or not self.use_flash_attn:
|
205 |
mixer_kwargs = {"adapter_mask": adapter_mask}
|
206 |
if key_padding_mask is not None:
|
207 |
mixer_kwargs["key_padding_mask"] = key_padding_mask.bool()
|
208 |
for layer in self.layers:
|
|
|
|
|
209 |
if self._grad_checkpointing:
|
210 |
hidden_states = torch.utils.checkpoint.checkpoint(
|
211 |
layer,
|
@@ -215,10 +228,14 @@ class XLMRobertaEncoder(nn.Module):
|
|
215 |
)
|
216 |
else:
|
217 |
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
|
|
|
|
218 |
if subset_mask is not None:
|
219 |
hidden_states = hidden_states[subset_mask]
|
220 |
else:
|
221 |
batch, seqlen = hidden_states.shape[:2]
|
|
|
|
|
222 |
hidden_states, indices, cu_seqlens, max_seqlen_in_batch, cu_adapter_mask = (
|
223 |
unpad_input(hidden_states, key_padding_mask, adapter_mask)
|
224 |
)
|
@@ -239,6 +256,10 @@ class XLMRobertaEncoder(nn.Module):
|
|
239 |
)
|
240 |
else:
|
241 |
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
|
|
|
|
|
|
|
|
242 |
hidden_states = pad_input(hidden_states, indices, batch, seqlen)
|
243 |
else:
|
244 |
for layer in self.layers[:-1]:
|
@@ -291,7 +312,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
291 |
hidden_states = self.layers[-1](
|
292 |
hidden_states_subset, mixer_kwargs=mixer_kwargs
|
293 |
)
|
294 |
-
return hidden_states
|
295 |
|
296 |
|
297 |
class XLMRobertaPooler(nn.Module):
|
@@ -588,7 +609,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
588 |
embeddings = self.mean_pooling(
|
589 |
token_embs, encoded_input["attention_mask"]
|
590 |
)
|
591 |
-
|
592 |
all_embeddings.extend(embeddings)
|
593 |
|
594 |
all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
|
@@ -596,9 +617,12 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
596 |
truncate_dim = truncate_dim or self.config.truncate_dim
|
597 |
if truncate_dim:
|
598 |
all_embeddings = self.truncate_embeddings(all_embeddings, truncate_dim)
|
599 |
-
|
600 |
if normalize_embeddings:
|
601 |
-
all_embeddings = [
|
|
|
|
|
|
|
602 |
|
603 |
if convert_to_tensor:
|
604 |
all_embeddings = torch.stack(all_embeddings)
|
@@ -659,6 +683,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
659 |
attention_mask=None,
|
660 |
masked_tokens_mask=None,
|
661 |
return_dict=None,
|
|
|
662 |
**kwargs,
|
663 |
):
|
664 |
"""If masked_tokens_mask is not None (i.e. last_layer_subset == True in XLMForPreTraining),
|
@@ -711,8 +736,15 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
711 |
key_padding_mask=attention_mask,
|
712 |
subset_mask=subset_mask,
|
713 |
adapter_mask=adapter_mask,
|
|
|
714 |
)
|
715 |
|
|
|
|
|
|
|
|
|
|
|
|
|
716 |
if masked_tokens_mask is None:
|
717 |
pooled_output = (
|
718 |
self.pooler(sequence_output, adapter_mask=adapter_mask)
|
@@ -742,6 +774,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
742 |
return BaseModelOutputWithPoolingAndCrossAttentions(
|
743 |
last_hidden_state=sequence_output,
|
744 |
pooler_output=pooled_output,
|
|
|
745 |
)
|
746 |
|
747 |
|
|
|
22 |
import torch.utils.checkpoint
|
23 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
24 |
from transformers import AutoTokenizer, PretrainedConfig
|
25 |
+
from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
|
|
|
26 |
from transformers.modeling_utils import PreTrainedModel
|
27 |
from transformers.models.bert.modeling_bert import (
|
28 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
29 |
+
BertForPreTrainingOutput,
|
30 |
+
)
|
31 |
+
from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
|
32 |
|
33 |
from .rotary import RotaryEmbedding
|
34 |
from .block import Block
|
|
|
195 |
self._grad_checkpointing = value
|
196 |
|
197 |
def forward(
|
198 |
+
self,
|
199 |
+
hidden_states,
|
200 |
+
key_padding_mask=None,
|
201 |
+
subset_mask=None,
|
202 |
+
adapter_mask=None,
|
203 |
+
output_hidden_states: Optional[bool] = None,
|
204 |
):
|
205 |
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
206 |
This means that we only compute the last layer output for these tokens.
|
207 |
subset_mask: (batch, seqlen), dtype=torch.bool
|
208 |
"""
|
209 |
+
|
210 |
+
all_hidden_states = () if output_hidden_states else None
|
211 |
+
|
212 |
+
if output_hidden_states and subset_mask:
|
213 |
+
raise ValueError('output_hidden_states is not supported for subset_masks')
|
214 |
+
|
215 |
if key_padding_mask is None or not self.use_flash_attn:
|
216 |
mixer_kwargs = {"adapter_mask": adapter_mask}
|
217 |
if key_padding_mask is not None:
|
218 |
mixer_kwargs["key_padding_mask"] = key_padding_mask.bool()
|
219 |
for layer in self.layers:
|
220 |
+
if output_hidden_states:
|
221 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
222 |
if self._grad_checkpointing:
|
223 |
hidden_states = torch.utils.checkpoint.checkpoint(
|
224 |
layer,
|
|
|
228 |
)
|
229 |
else:
|
230 |
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
231 |
+
if output_hidden_states:
|
232 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
233 |
if subset_mask is not None:
|
234 |
hidden_states = hidden_states[subset_mask]
|
235 |
else:
|
236 |
batch, seqlen = hidden_states.shape[:2]
|
237 |
+
if output_hidden_states:
|
238 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
239 |
hidden_states, indices, cu_seqlens, max_seqlen_in_batch, cu_adapter_mask = (
|
240 |
unpad_input(hidden_states, key_padding_mask, adapter_mask)
|
241 |
)
|
|
|
256 |
)
|
257 |
else:
|
258 |
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
259 |
+
if output_hidden_states:
|
260 |
+
all_hidden_states = all_hidden_states + (
|
261 |
+
pad_input(hidden_states, indices, batch, seqlen),
|
262 |
+
)
|
263 |
hidden_states = pad_input(hidden_states, indices, batch, seqlen)
|
264 |
else:
|
265 |
for layer in self.layers[:-1]:
|
|
|
312 |
hidden_states = self.layers[-1](
|
313 |
hidden_states_subset, mixer_kwargs=mixer_kwargs
|
314 |
)
|
315 |
+
return all_hidden_states if output_hidden_states else hidden_states
|
316 |
|
317 |
|
318 |
class XLMRobertaPooler(nn.Module):
|
|
|
609 |
embeddings = self.mean_pooling(
|
610 |
token_embs, encoded_input["attention_mask"]
|
611 |
)
|
612 |
+
|
613 |
all_embeddings.extend(embeddings)
|
614 |
|
615 |
all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
|
|
|
617 |
truncate_dim = truncate_dim or self.config.truncate_dim
|
618 |
if truncate_dim:
|
619 |
all_embeddings = self.truncate_embeddings(all_embeddings, truncate_dim)
|
620 |
+
|
621 |
if normalize_embeddings:
|
622 |
+
all_embeddings = [
|
623 |
+
torch.nn.functional.normalize(embedding, p=2, dim=0)
|
624 |
+
for embedding in all_embeddings
|
625 |
+
]
|
626 |
|
627 |
if convert_to_tensor:
|
628 |
all_embeddings = torch.stack(all_embeddings)
|
|
|
683 |
attention_mask=None,
|
684 |
masked_tokens_mask=None,
|
685 |
return_dict=None,
|
686 |
+
output_hidden_states=None,
|
687 |
**kwargs,
|
688 |
):
|
689 |
"""If masked_tokens_mask is not None (i.e. last_layer_subset == True in XLMForPreTraining),
|
|
|
736 |
key_padding_mask=attention_mask,
|
737 |
subset_mask=subset_mask,
|
738 |
adapter_mask=adapter_mask,
|
739 |
+
output_hidden_states=output_hidden_states,
|
740 |
)
|
741 |
|
742 |
+
if output_hidden_states:
|
743 |
+
all_hidden_states = sequence_output
|
744 |
+
sequence_output = sequence_output[-1]
|
745 |
+
else:
|
746 |
+
all_hidden_states = None
|
747 |
+
|
748 |
if masked_tokens_mask is None:
|
749 |
pooled_output = (
|
750 |
self.pooler(sequence_output, adapter_mask=adapter_mask)
|
|
|
774 |
return BaseModelOutputWithPoolingAndCrossAttentions(
|
775 |
last_hidden_state=sequence_output,
|
776 |
pooler_output=pooled_output,
|
777 |
+
hidden_states=all_hidden_states,
|
778 |
)
|
779 |
|
780 |
|