Spaces:
Running
Running
mohsenfayyaz
commited on
Commit
•
e654c3a
1
Parent(s):
f34a8cd
Upload 3 files
Browse files- DecompX/src/decompx_utils.py +50 -0
- DecompX/src/modeling_bert.py +100 -100
- DecompX/src/modeling_roberta.py +300 -285
DecompX/src/decompx_utils.py
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from dataclasses import dataclass
|
3 |
+
from typing import List, Optional, Tuple, Union
|
4 |
+
|
5 |
+
|
6 |
+
@dataclass
|
7 |
+
class DecompXConfig():
|
8 |
+
include_biases: Optional[bool] = True
|
9 |
+
bias_decomp_type: Optional[str] = "absdot" # "absdot": Based on the absolute value of dot products | "norm": Based on the norm of the attribution vectors | "equal": equal decomposition | "abssim": Based on the absolute value of cosine similarites | "cls": add to cls token
|
10 |
+
include_bias_token: Optional[bool] = False # Adds an extra input token as a bias in the attribution vectors
|
11 |
+
# If the bias_decomp_type is None and include_bias_token is True then the final token in the input tokens of the attr. vectors will be the summation of the biases
|
12 |
+
# Otherwise the bias token will be decomposed with the specified decomp type
|
13 |
+
|
14 |
+
include_LN1: Optional[bool] = True
|
15 |
+
|
16 |
+
include_FFN: Optional[bool] = True
|
17 |
+
FFN_approx_type: Optional[str] = "GeLU_ZO" # "GeLU_LA": GeLU-based linear approximation | "ReLU": Using ReLU as an approximation | "GeLU_ZO": Zero-origin slope approximation
|
18 |
+
FFN_fast_mode: Optional[bool] = False
|
19 |
+
|
20 |
+
include_LN2: Optional[bool] = True
|
21 |
+
|
22 |
+
aggregation: Optional[str] = None # None: No aggregation | vector: Vector-based aggregation | rollout: Norm-based rollout aggregation
|
23 |
+
|
24 |
+
include_classifier_w_pooler: Optional[bool] = True
|
25 |
+
tanh_approx_type: Optional[str] = "ZO" # "ZO": Zero-origin slope approximation | "LA": Linear approximation
|
26 |
+
|
27 |
+
output_all_layers: Optional[bool] = False # True: Output all layers | False: Output only last layer
|
28 |
+
output_attention: Optional[str] = None # None | norm | vector | both
|
29 |
+
output_res1: Optional[str] = None # None | norm | vector | both
|
30 |
+
output_LN1: Optional[str] = None # None | norm | vector | both
|
31 |
+
output_FFN: Optional[str] = None # None | norm | vector | both
|
32 |
+
output_res2: Optional[str] = None # None | norm | vector | both
|
33 |
+
output_encoder: Optional[str] = None # None | norm | vector | both
|
34 |
+
output_aggregated: Optional[str] = None # None | norm | vector | both
|
35 |
+
output_pooler: Optional[str] = None # None | norm | vector | both
|
36 |
+
|
37 |
+
output_classifier: Optional[bool] = True
|
38 |
+
|
39 |
+
|
40 |
+
@dataclass
|
41 |
+
class DecompXOutput():
|
42 |
+
attention: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
43 |
+
res1: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
44 |
+
LN1: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
45 |
+
FFN: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
46 |
+
res2: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
47 |
+
encoder: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
48 |
+
aggregated: Optional[Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Tuple[torch.Tensor], torch.Tensor]] = None
|
49 |
+
pooler: Optional[Union[Tuple[torch.Tensor], torch.Tensor]] = None
|
50 |
+
classifier: Optional[torch.Tensor] = None
|
DecompX/src/modeling_bert.py
CHANGED
@@ -27,7 +27,7 @@ from packaging import version
|
|
27 |
from torch import nn
|
28 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
29 |
|
30 |
-
from .
|
31 |
|
32 |
from transformers.activations import ACT2FN
|
33 |
from transformers.modeling_outputs import (
|
@@ -289,7 +289,7 @@ class BertSelfAttention(nn.Module):
|
|
289 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
290 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
291 |
output_attentions: Optional[bool] = False,
|
292 |
-
|
293 |
) -> Tuple[torch.Tensor]:
|
294 |
mixed_query_layer = self.query(hidden_states)
|
295 |
|
@@ -376,7 +376,7 @@ class BertSelfAttention(nn.Module):
|
|
376 |
|
377 |
# added by Fayyaz / Modarressi
|
378 |
# -------------------------------
|
379 |
-
if
|
380 |
outputs = (context_layer, attention_probs, value_layer, decomposed_value_layer)
|
381 |
return outputs
|
382 |
# -------------------------------
|
@@ -396,14 +396,14 @@ class BertSelfOutput(nn.Module):
|
|
396 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
397 |
|
398 |
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
399 |
-
|
400 |
hidden_states = self.dense(hidden_states)
|
401 |
hidden_states = self.dropout(hidden_states)
|
402 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
403 |
pre_ln_states = hidden_states + input_tensor # added by Fayyaz / Modarressi
|
404 |
post_ln_states = self.LayerNorm(pre_ln_states) # added by Fayyaz / Modarressi
|
405 |
# added by Fayyaz / Modarressi
|
406 |
-
if
|
407 |
return post_ln_states, pre_ln_states
|
408 |
else:
|
409 |
return post_ln_states
|
@@ -444,7 +444,7 @@ class BertAttention(nn.Module):
|
|
444 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
445 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
446 |
output_attentions: Optional[bool] = False,
|
447 |
-
|
448 |
) -> Tuple[torch.Tensor]:
|
449 |
self_outputs = self.self(
|
450 |
hidden_states,
|
@@ -455,17 +455,17 @@ class BertAttention(nn.Module):
|
|
455 |
encoder_attention_mask,
|
456 |
past_key_value,
|
457 |
output_attentions,
|
458 |
-
|
459 |
)
|
460 |
attention_output = self.output(
|
461 |
self_outputs[0],
|
462 |
hidden_states,
|
463 |
-
|
464 |
)
|
465 |
|
466 |
# Added by Fayyaz / Modarressi
|
467 |
# -------------------------------
|
468 |
-
if
|
469 |
_, attention_probs, value_layer, decomposed_value_layer = self_outputs
|
470 |
attention_output, pre_ln_states = attention_output
|
471 |
outputs = (attention_output, attention_probs,) + (value_layer, decomposed_value_layer, pre_ln_states) # add attentions and norms if we output them
|
@@ -485,10 +485,10 @@ class BertIntermediate(nn.Module):
|
|
485 |
else:
|
486 |
self.intermediate_act_fn = config.hidden_act
|
487 |
|
488 |
-
def forward(self, hidden_states: torch.Tensor,
|
489 |
pre_act_hidden_states = self.dense(hidden_states)
|
490 |
hidden_states = self.intermediate_act_fn(pre_act_hidden_states)
|
491 |
-
if
|
492 |
return hidden_states, pre_act_hidden_states
|
493 |
return hidden_states, None
|
494 |
|
@@ -500,7 +500,7 @@ class BertOutput(nn.Module):
|
|
500 |
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
501 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
502 |
|
503 |
-
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
504 |
hidden_states = self.dense(hidden_states)
|
505 |
hidden_states = self.dropout(hidden_states)
|
506 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
@@ -509,7 +509,7 @@ class BertOutput(nn.Module):
|
|
509 |
# -------------------------------
|
510 |
pre_ln_states = hidden_states + input_tensor
|
511 |
hidden_states = self.LayerNorm(pre_ln_states)
|
512 |
-
if
|
513 |
return hidden_states, pre_ln_states
|
514 |
return hidden_states, None
|
515 |
# -------------------------------
|
@@ -689,7 +689,7 @@ class BertLayer(nn.Module):
|
|
689 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
690 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
691 |
output_attentions: Optional[bool] = False,
|
692 |
-
|
693 |
) -> Tuple[torch.Tensor]:
|
694 |
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
695 |
# self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
@@ -700,14 +700,14 @@ class BertLayer(nn.Module):
|
|
700 |
# output_attentions=output_attentions,
|
701 |
# past_key_value=self_attn_past_key_value,
|
702 |
# )
|
703 |
-
|
704 |
self_attention_outputs = self.attention(
|
705 |
hidden_states,
|
706 |
attribution_vectors,
|
707 |
attention_mask,
|
708 |
head_mask,
|
709 |
output_attentions=output_attentions,
|
710 |
-
|
711 |
) # changed by Goro Kobayashi
|
712 |
attention_output = self_attention_outputs[0]
|
713 |
|
@@ -749,16 +749,16 @@ class BertLayer(nn.Module):
|
|
749 |
|
750 |
# Added by Fayyaz / Modarressi
|
751 |
# -------------------------------
|
752 |
-
bias_decomp_type = "biastoken" if
|
753 |
-
intermediate_output, pre_act_hidden_states = self.intermediate(attention_output,
|
754 |
-
layer_output, pre_ln2_states = self.output(intermediate_output, attention_output,
|
755 |
-
if
|
756 |
attention_probs, value_layer, decomposed_value_layer, pre_ln_states = outputs
|
757 |
|
758 |
headmixing_weight = self.attention.output.dense.weight.view(self.all_head_size, self.num_attention_heads,
|
759 |
self.attention_head_size)
|
760 |
|
761 |
-
if decomposed_value_layer is None or
|
762 |
transformed_layer = torch.einsum('bhsv,dhv->bhsd', value_layer, headmixing_weight) # V * W^o (z=(qk)v)
|
763 |
# Make weighted vectors αf(x) from transformed vectors (transformed_layer)
|
764 |
# and attention weights (attentions):
|
@@ -789,29 +789,29 @@ class BertLayer(nn.Module):
|
|
789 |
residual_weighted_layer = summed_weighted_layer + attribution_vectors
|
790 |
accumulated_bias = torch.matmul(self.attention.output.dense.weight, self.attention.self.value.bias) + self.attention.output.dense.bias
|
791 |
|
792 |
-
if
|
793 |
residual_weighted_layer = self.bias_decomposer(accumulated_bias, residual_weighted_layer, bias_decomp_type)
|
794 |
|
795 |
-
if
|
796 |
post_ln_layer = self.ln_decomposer(
|
797 |
attribution_vectors=residual_weighted_layer,
|
798 |
pre_ln_states=pre_ln_states,
|
799 |
gamma=self.attention.output.LayerNorm.weight.data,
|
800 |
beta=self.attention.output.LayerNorm.bias.data,
|
801 |
eps=self.attention.output.LayerNorm.eps,
|
802 |
-
include_biases=
|
803 |
bias_decomp_type=bias_decomp_type
|
804 |
)
|
805 |
else:
|
806 |
post_ln_layer = residual_weighted_layer
|
807 |
|
808 |
-
if
|
809 |
-
post_ffn_layer = self.ffn_decomposer_fast if
|
810 |
attribution_vectors=post_ln_layer,
|
811 |
intermediate_hidden_states=pre_act_hidden_states,
|
812 |
intermediate_output=intermediate_output,
|
813 |
-
approximation_type=
|
814 |
-
include_biases=
|
815 |
bias_decomp_type=bias_decomp_type
|
816 |
)
|
817 |
pre_ln2_layer = post_ln_layer + post_ffn_layer
|
@@ -819,25 +819,25 @@ class BertLayer(nn.Module):
|
|
819 |
pre_ln2_layer = post_ln_layer
|
820 |
post_ffn_layer = None
|
821 |
|
822 |
-
if
|
823 |
post_ln2_layer = self.ln_decomposer(
|
824 |
attribution_vectors=pre_ln2_layer,
|
825 |
pre_ln_states=pre_ln2_states,
|
826 |
gamma=self.output.LayerNorm.weight.data,
|
827 |
beta=self.output.LayerNorm.bias.data,
|
828 |
eps=self.output.LayerNorm.eps,
|
829 |
-
include_biases=
|
830 |
bias_decomp_type=bias_decomp_type
|
831 |
)
|
832 |
else:
|
833 |
post_ln2_layer = pre_ln2_layer
|
834 |
|
835 |
-
new_outputs =
|
836 |
-
attention=output_builder(summed_weighted_layer,
|
837 |
-
res1=output_builder(residual_weighted_layer,
|
838 |
-
LN1=output_builder(post_ln_layer,
|
839 |
-
FFN=output_builder(post_ffn_layer,
|
840 |
-
res2=output_builder(pre_ln2_layer,
|
841 |
encoder=output_builder(post_ln2_layer, "both")
|
842 |
)
|
843 |
return (layer_output,) + (new_outputs,)
|
@@ -875,7 +875,7 @@ class BertEncoder(nn.Module):
|
|
875 |
output_attentions: Optional[bool] = False,
|
876 |
output_hidden_states: Optional[bool] = False,
|
877 |
return_dict: Optional[bool] = True,
|
878 |
-
|
879 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
|
880 |
all_hidden_states = () if output_hidden_states else None
|
881 |
all_self_attentions = () if output_attentions else None
|
@@ -887,18 +887,18 @@ class BertEncoder(nn.Module):
|
|
887 |
aggregated_encoder_vectors = None # added by Fayyaz / Modarressi
|
888 |
|
889 |
# -- added by Fayyaz / Modarressi
|
890 |
-
if
|
891 |
-
|
892 |
-
attention=() if
|
893 |
-
res1=() if
|
894 |
-
LN1=() if
|
895 |
-
FFN=() if
|
896 |
-
res2=() if
|
897 |
-
encoder=() if
|
898 |
-
aggregated=() if
|
899 |
)
|
900 |
else:
|
901 |
-
|
902 |
# -- added by Fayyaz / Modarressi
|
903 |
|
904 |
for i, layer_module in enumerate(self.layer):
|
@@ -940,7 +940,7 @@ class BertEncoder(nn.Module):
|
|
940 |
encoder_attention_mask,
|
941 |
past_key_value,
|
942 |
output_attentions,
|
943 |
-
|
944 |
)
|
945 |
|
946 |
hidden_states = layer_outputs[0]
|
@@ -952,47 +952,47 @@ class BertEncoder(nn.Module):
|
|
952 |
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
|
953 |
|
954 |
# added by Fayyaz / Modarressi
|
955 |
-
if
|
956 |
-
|
957 |
-
if
|
958 |
-
if
|
959 |
raise Exception("Classifier and pooler could be included in vector aggregation mode")
|
960 |
|
961 |
-
encoder_norms =
|
962 |
|
963 |
if aggregated_encoder_norms is None:
|
964 |
aggregated_encoder_norms = encoder_norms * torch.exp(attention_mask).view((-1, attention_mask.shape[-1], 1))
|
965 |
else:
|
966 |
aggregated_encoder_norms = torch.einsum("ijk,ikm->ijm", encoder_norms, aggregated_encoder_norms)
|
967 |
|
968 |
-
if
|
969 |
-
|
970 |
-
elif
|
971 |
raise Exception("Rollout aggregated values are only available in norms. Set output_aggregated to 'norm'.")
|
972 |
|
973 |
|
974 |
-
elif
|
975 |
-
aggregated_encoder_vectors =
|
976 |
|
977 |
-
if
|
978 |
-
|
979 |
else:
|
980 |
-
|
981 |
|
982 |
-
|
983 |
|
984 |
-
if
|
985 |
-
|
986 |
-
|
987 |
-
|
988 |
-
|
989 |
-
|
990 |
-
|
991 |
|
992 |
-
if
|
993 |
-
|
994 |
else:
|
995 |
-
|
996 |
|
997 |
if output_hidden_states:
|
998 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
@@ -1006,8 +1006,8 @@ class BertEncoder(nn.Module):
|
|
1006 |
all_hidden_states,
|
1007 |
all_self_attentions,
|
1008 |
all_cross_attentions,
|
1009 |
-
|
1010 |
-
|
1011 |
]
|
1012 |
if v is not None
|
1013 |
)
|
@@ -1026,13 +1026,13 @@ class BertPooler(nn.Module):
|
|
1026 |
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
1027 |
self.activation = nn.Tanh()
|
1028 |
|
1029 |
-
def forward(self, hidden_states: torch.Tensor,
|
1030 |
# We "pool" the model by simply taking the hidden state corresponding
|
1031 |
# to the first token.
|
1032 |
first_token_tensor = hidden_states[:, 0]
|
1033 |
pre_pooled_output = self.dense(first_token_tensor)
|
1034 |
pooled_output = self.activation(pre_pooled_output)
|
1035 |
-
if
|
1036 |
return pooled_output, pre_pooled_output
|
1037 |
return pooled_output
|
1038 |
|
@@ -1378,7 +1378,7 @@ class BertModel(BertPreTrainedModel):
|
|
1378 |
output_attentions: Optional[bool] = None,
|
1379 |
output_hidden_states: Optional[bool] = None,
|
1380 |
return_dict: Optional[bool] = None,
|
1381 |
-
|
1382 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
|
1383 |
r"""
|
1384 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
@@ -1477,32 +1477,32 @@ class BertModel(BertPreTrainedModel):
|
|
1477 |
output_attentions=output_attentions,
|
1478 |
output_hidden_states=output_hidden_states,
|
1479 |
return_dict=return_dict,
|
1480 |
-
|
1481 |
)
|
1482 |
sequence_output = encoder_outputs[0]
|
1483 |
-
|
1484 |
-
pooled_output = self.pooler(sequence_output,
|
1485 |
|
1486 |
-
if
|
1487 |
pre_act_pooled = pooled_output[1]
|
1488 |
pooled_output = pooled_output[0]
|
1489 |
|
1490 |
-
if
|
1491 |
-
|
1492 |
-
aggregated_attribution_vectors = encoder_outputs[
|
1493 |
|
1494 |
-
encoder_outputs[
|
1495 |
|
1496 |
pooler_decomposed = self.ffn_decomposer(
|
1497 |
attribution_vectors=aggregated_attribution_vectors[:, 0],
|
1498 |
pre_act_pooled=pre_act_pooled,
|
1499 |
post_act_pooled=pooled_output,
|
1500 |
-
include_biases=
|
1501 |
-
bias_decomp_type="biastoken" if
|
1502 |
-
tanh_approx_type=
|
1503 |
)
|
1504 |
|
1505 |
-
encoder_outputs[
|
1506 |
|
1507 |
if not return_dict:
|
1508 |
return (sequence_output, pooled_output) + encoder_outputs[1:]
|
@@ -2085,7 +2085,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
|
|
2085 |
output_attentions: Optional[bool] = None,
|
2086 |
output_hidden_states: Optional[bool] = None,
|
2087 |
return_dict: Optional[bool] = None,
|
2088 |
-
|
2089 |
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
|
2090 |
r"""
|
2091 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
@@ -2105,7 +2105,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
|
|
2105 |
output_attentions=output_attentions,
|
2106 |
output_hidden_states=output_hidden_states,
|
2107 |
return_dict=return_dict,
|
2108 |
-
|
2109 |
)
|
2110 |
|
2111 |
pooled_output = outputs[1]
|
@@ -2113,29 +2113,29 @@ class BertForSequenceClassification(BertPreTrainedModel):
|
|
2113 |
pooled_output = self.dropout(pooled_output)
|
2114 |
logits = self.classifier(pooled_output)
|
2115 |
|
2116 |
-
if
|
2117 |
-
|
2118 |
-
aggregated_attribution_vectors = outputs[
|
2119 |
|
2120 |
-
outputs[
|
2121 |
|
2122 |
classifier_decomposed = self.ffn_decomposer(
|
2123 |
attribution_vectors=aggregated_attribution_vectors,
|
2124 |
-
include_biases=
|
2125 |
-
bias_decomp_type="biastoken" if
|
2126 |
)
|
2127 |
|
2128 |
-
if
|
2129 |
bias_token = classifier_decomposed[:,-1,:].detach().clone()
|
2130 |
classifier_decomposed = classifier_decomposed[:,:-1,:]
|
2131 |
classifier_decomposed = self.biastoken_decomposer(
|
2132 |
bias_token,
|
2133 |
classifier_decomposed,
|
2134 |
-
bias_decomp_type=
|
2135 |
)
|
2136 |
|
2137 |
|
2138 |
-
outputs[
|
2139 |
|
2140 |
loss = None
|
2141 |
if labels is not None:
|
|
|
27 |
from torch import nn
|
28 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
29 |
|
30 |
+
from .decompx_utils import DecompXConfig, DecompXOutput
|
31 |
|
32 |
from transformers.activations import ACT2FN
|
33 |
from transformers.modeling_outputs import (
|
|
|
289 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
290 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
291 |
output_attentions: Optional[bool] = False,
|
292 |
+
decompx_ready: Optional[bool] = None, # added by Fayyaz / Modarressi
|
293 |
) -> Tuple[torch.Tensor]:
|
294 |
mixed_query_layer = self.query(hidden_states)
|
295 |
|
|
|
376 |
|
377 |
# added by Fayyaz / Modarressi
|
378 |
# -------------------------------
|
379 |
+
if decompx_ready:
|
380 |
outputs = (context_layer, attention_probs, value_layer, decomposed_value_layer)
|
381 |
return outputs
|
382 |
# -------------------------------
|
|
|
396 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
397 |
|
398 |
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
399 |
+
decompx_ready=False): # added by Fayyaz / Modarressi
|
400 |
hidden_states = self.dense(hidden_states)
|
401 |
hidden_states = self.dropout(hidden_states)
|
402 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
403 |
pre_ln_states = hidden_states + input_tensor # added by Fayyaz / Modarressi
|
404 |
post_ln_states = self.LayerNorm(pre_ln_states) # added by Fayyaz / Modarressi
|
405 |
# added by Fayyaz / Modarressi
|
406 |
+
if decompx_ready:
|
407 |
return post_ln_states, pre_ln_states
|
408 |
else:
|
409 |
return post_ln_states
|
|
|
444 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
445 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
446 |
output_attentions: Optional[bool] = False,
|
447 |
+
decompx_ready: Optional[bool] = None, # added by Fayyaz / Modarressi
|
448 |
) -> Tuple[torch.Tensor]:
|
449 |
self_outputs = self.self(
|
450 |
hidden_states,
|
|
|
455 |
encoder_attention_mask,
|
456 |
past_key_value,
|
457 |
output_attentions,
|
458 |
+
decompx_ready=decompx_ready, # added by Fayyaz / Modarressi
|
459 |
)
|
460 |
attention_output = self.output(
|
461 |
self_outputs[0],
|
462 |
hidden_states,
|
463 |
+
decompx_ready=decompx_ready, # added by Goro Kobayashi (Edited by Fayyaz / Modarressi)
|
464 |
)
|
465 |
|
466 |
# Added by Fayyaz / Modarressi
|
467 |
# -------------------------------
|
468 |
+
if decompx_ready:
|
469 |
_, attention_probs, value_layer, decomposed_value_layer = self_outputs
|
470 |
attention_output, pre_ln_states = attention_output
|
471 |
outputs = (attention_output, attention_probs,) + (value_layer, decomposed_value_layer, pre_ln_states) # add attentions and norms if we output them
|
|
|
485 |
else:
|
486 |
self.intermediate_act_fn = config.hidden_act
|
487 |
|
488 |
+
def forward(self, hidden_states: torch.Tensor, decompx_ready: Optional[bool] = False) -> torch.Tensor:
|
489 |
pre_act_hidden_states = self.dense(hidden_states)
|
490 |
hidden_states = self.intermediate_act_fn(pre_act_hidden_states)
|
491 |
+
if decompx_ready:
|
492 |
return hidden_states, pre_act_hidden_states
|
493 |
return hidden_states, None
|
494 |
|
|
|
500 |
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
501 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
502 |
|
503 |
+
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor, decompx_ready: Optional[bool] = False):
|
504 |
hidden_states = self.dense(hidden_states)
|
505 |
hidden_states = self.dropout(hidden_states)
|
506 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
|
|
509 |
# -------------------------------
|
510 |
pre_ln_states = hidden_states + input_tensor
|
511 |
hidden_states = self.LayerNorm(pre_ln_states)
|
512 |
+
if decompx_ready:
|
513 |
return hidden_states, pre_ln_states
|
514 |
return hidden_states, None
|
515 |
# -------------------------------
|
|
|
689 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
690 |
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
691 |
output_attentions: Optional[bool] = False,
|
692 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
693 |
) -> Tuple[torch.Tensor]:
|
694 |
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
695 |
# self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
|
|
700 |
# output_attentions=output_attentions,
|
701 |
# past_key_value=self_attn_past_key_value,
|
702 |
# )
|
703 |
+
decompx_ready = decompx_config is not None
|
704 |
self_attention_outputs = self.attention(
|
705 |
hidden_states,
|
706 |
attribution_vectors,
|
707 |
attention_mask,
|
708 |
head_mask,
|
709 |
output_attentions=output_attentions,
|
710 |
+
decompx_ready=decompx_ready,
|
711 |
) # changed by Goro Kobayashi
|
712 |
attention_output = self_attention_outputs[0]
|
713 |
|
|
|
749 |
|
750 |
# Added by Fayyaz / Modarressi
|
751 |
# -------------------------------
|
752 |
+
bias_decomp_type = "biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type
|
753 |
+
intermediate_output, pre_act_hidden_states = self.intermediate(attention_output, decompx_ready=decompx_ready)
|
754 |
+
layer_output, pre_ln2_states = self.output(intermediate_output, attention_output, decompx_ready=decompx_ready)
|
755 |
+
if decompx_ready:
|
756 |
attention_probs, value_layer, decomposed_value_layer, pre_ln_states = outputs
|
757 |
|
758 |
headmixing_weight = self.attention.output.dense.weight.view(self.all_head_size, self.num_attention_heads,
|
759 |
self.attention_head_size)
|
760 |
|
761 |
+
if decomposed_value_layer is None or decompx_config.aggregation != "vector":
|
762 |
transformed_layer = torch.einsum('bhsv,dhv->bhsd', value_layer, headmixing_weight) # V * W^o (z=(qk)v)
|
763 |
# Make weighted vectors αf(x) from transformed vectors (transformed_layer)
|
764 |
# and attention weights (attentions):
|
|
|
789 |
residual_weighted_layer = summed_weighted_layer + attribution_vectors
|
790 |
accumulated_bias = torch.matmul(self.attention.output.dense.weight, self.attention.self.value.bias) + self.attention.output.dense.bias
|
791 |
|
792 |
+
if decompx_config.include_biases:
|
793 |
residual_weighted_layer = self.bias_decomposer(accumulated_bias, residual_weighted_layer, bias_decomp_type)
|
794 |
|
795 |
+
if decompx_config.include_LN1:
|
796 |
post_ln_layer = self.ln_decomposer(
|
797 |
attribution_vectors=residual_weighted_layer,
|
798 |
pre_ln_states=pre_ln_states,
|
799 |
gamma=self.attention.output.LayerNorm.weight.data,
|
800 |
beta=self.attention.output.LayerNorm.bias.data,
|
801 |
eps=self.attention.output.LayerNorm.eps,
|
802 |
+
include_biases=decompx_config.include_biases,
|
803 |
bias_decomp_type=bias_decomp_type
|
804 |
)
|
805 |
else:
|
806 |
post_ln_layer = residual_weighted_layer
|
807 |
|
808 |
+
if decompx_config.include_FFN:
|
809 |
+
post_ffn_layer = self.ffn_decomposer_fast if decompx_config.FFN_fast_mode else self.ffn_decomposer(
|
810 |
attribution_vectors=post_ln_layer,
|
811 |
intermediate_hidden_states=pre_act_hidden_states,
|
812 |
intermediate_output=intermediate_output,
|
813 |
+
approximation_type=decompx_config.FFN_approx_type,
|
814 |
+
include_biases=decompx_config.include_biases,
|
815 |
bias_decomp_type=bias_decomp_type
|
816 |
)
|
817 |
pre_ln2_layer = post_ln_layer + post_ffn_layer
|
|
|
819 |
pre_ln2_layer = post_ln_layer
|
820 |
post_ffn_layer = None
|
821 |
|
822 |
+
if decompx_config.include_LN2:
|
823 |
post_ln2_layer = self.ln_decomposer(
|
824 |
attribution_vectors=pre_ln2_layer,
|
825 |
pre_ln_states=pre_ln2_states,
|
826 |
gamma=self.output.LayerNorm.weight.data,
|
827 |
beta=self.output.LayerNorm.bias.data,
|
828 |
eps=self.output.LayerNorm.eps,
|
829 |
+
include_biases=decompx_config.include_biases,
|
830 |
bias_decomp_type=bias_decomp_type
|
831 |
)
|
832 |
else:
|
833 |
post_ln2_layer = pre_ln2_layer
|
834 |
|
835 |
+
new_outputs = DecompXOutput(
|
836 |
+
attention=output_builder(summed_weighted_layer, decompx_config.output_attention),
|
837 |
+
res1=output_builder(residual_weighted_layer, decompx_config.output_res1),
|
838 |
+
LN1=output_builder(post_ln_layer, decompx_config.output_res2),
|
839 |
+
FFN=output_builder(post_ffn_layer, decompx_config.output_FFN),
|
840 |
+
res2=output_builder(pre_ln2_layer, decompx_config.output_res2),
|
841 |
encoder=output_builder(post_ln2_layer, "both")
|
842 |
)
|
843 |
return (layer_output,) + (new_outputs,)
|
|
|
875 |
output_attentions: Optional[bool] = False,
|
876 |
output_hidden_states: Optional[bool] = False,
|
877 |
return_dict: Optional[bool] = True,
|
878 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
879 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
|
880 |
all_hidden_states = () if output_hidden_states else None
|
881 |
all_self_attentions = () if output_attentions else None
|
|
|
887 |
aggregated_encoder_vectors = None # added by Fayyaz / Modarressi
|
888 |
|
889 |
# -- added by Fayyaz / Modarressi
|
890 |
+
if decompx_config and decompx_config.output_all_layers:
|
891 |
+
all_decompx_outputs = DecompXOutput(
|
892 |
+
attention=() if decompx_config.output_attention else None,
|
893 |
+
res1=() if decompx_config.output_res1 else None,
|
894 |
+
LN1=() if decompx_config.output_LN1 else None,
|
895 |
+
FFN=() if decompx_config.output_LN1 else None,
|
896 |
+
res2=() if decompx_config.output_res2 else None,
|
897 |
+
encoder=() if decompx_config.output_encoder else None,
|
898 |
+
aggregated=() if decompx_config.output_aggregated and decompx_config.aggregation else None,
|
899 |
)
|
900 |
else:
|
901 |
+
all_decompx_outputs = None
|
902 |
# -- added by Fayyaz / Modarressi
|
903 |
|
904 |
for i, layer_module in enumerate(self.layer):
|
|
|
940 |
encoder_attention_mask,
|
941 |
past_key_value,
|
942 |
output_attentions,
|
943 |
+
decompx_config # added by Fayyaz / Modarressi
|
944 |
)
|
945 |
|
946 |
hidden_states = layer_outputs[0]
|
|
|
952 |
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
|
953 |
|
954 |
# added by Fayyaz / Modarressi
|
955 |
+
if decompx_config:
|
956 |
+
decompx_output = layer_outputs[1]
|
957 |
+
if decompx_config.aggregation == "rollout":
|
958 |
+
if decompx_config.include_classifier_w_pooler:
|
959 |
raise Exception("Classifier and pooler could be included in vector aggregation mode")
|
960 |
|
961 |
+
encoder_norms = decompx_output.encoder[0][0]
|
962 |
|
963 |
if aggregated_encoder_norms is None:
|
964 |
aggregated_encoder_norms = encoder_norms * torch.exp(attention_mask).view((-1, attention_mask.shape[-1], 1))
|
965 |
else:
|
966 |
aggregated_encoder_norms = torch.einsum("ijk,ikm->ijm", encoder_norms, aggregated_encoder_norms)
|
967 |
|
968 |
+
if decompx_config.output_aggregated == "norm":
|
969 |
+
decompx_output.aggregated = (aggregated_encoder_norms,)
|
970 |
+
elif decompx_config.output_aggregated is not None:
|
971 |
raise Exception("Rollout aggregated values are only available in norms. Set output_aggregated to 'norm'.")
|
972 |
|
973 |
|
974 |
+
elif decompx_config.aggregation == "vector":
|
975 |
+
aggregated_encoder_vectors = decompx_output.encoder[0][1]
|
976 |
|
977 |
+
if decompx_config.include_classifier_w_pooler:
|
978 |
+
decompx_output.aggregated = (aggregated_encoder_vectors,)
|
979 |
else:
|
980 |
+
decompx_output.aggregated = output_builder(aggregated_encoder_vectors, decompx_config.output_aggregated)
|
981 |
|
982 |
+
decompx_output.encoder = output_builder(decompx_output.encoder[0][1], decompx_config.output_encoder)
|
983 |
|
984 |
+
if decompx_config.output_all_layers:
|
985 |
+
all_decompx_outputs.attention = all_decompx_outputs.attention + decompx_output.attention if decompx_config.output_attention else None
|
986 |
+
all_decompx_outputs.res1 = all_decompx_outputs.res1 + decompx_output.res1 if decompx_config.output_res1 else None
|
987 |
+
all_decompx_outputs.LN1 = all_decompx_outputs.LN1 + decompx_output.LN1 if decompx_config.output_LN1 else None
|
988 |
+
all_decompx_outputs.FFN = all_decompx_outputs.FFN + decompx_output.FFN if decompx_config.output_FFN else None
|
989 |
+
all_decompx_outputs.res2 = all_decompx_outputs.res2 + decompx_output.res2 if decompx_config.output_res2 else None
|
990 |
+
all_decompx_outputs.encoder = all_decompx_outputs.encoder + decompx_output.encoder if decompx_config.output_encoder else None
|
991 |
|
992 |
+
if decompx_config.include_classifier_w_pooler and decompx_config.aggregation == "vector":
|
993 |
+
all_decompx_outputs.aggregated = all_decompx_outputs.aggregated + output_builder(aggregated_encoder_vectors, decompx_config.output_aggregated) if decompx_config.output_aggregated else None
|
994 |
else:
|
995 |
+
all_decompx_outputs.aggregated = all_decompx_outputs.aggregated + decompx_output.aggregated if decompx_config.output_aggregated else None
|
996 |
|
997 |
if output_hidden_states:
|
998 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
1006 |
all_hidden_states,
|
1007 |
all_self_attentions,
|
1008 |
all_cross_attentions,
|
1009 |
+
decompx_output if decompx_config else None,
|
1010 |
+
all_decompx_outputs
|
1011 |
]
|
1012 |
if v is not None
|
1013 |
)
|
|
|
1026 |
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
1027 |
self.activation = nn.Tanh()
|
1028 |
|
1029 |
+
def forward(self, hidden_states: torch.Tensor, decompx_ready=False) -> torch.Tensor:
|
1030 |
# We "pool" the model by simply taking the hidden state corresponding
|
1031 |
# to the first token.
|
1032 |
first_token_tensor = hidden_states[:, 0]
|
1033 |
pre_pooled_output = self.dense(first_token_tensor)
|
1034 |
pooled_output = self.activation(pre_pooled_output)
|
1035 |
+
if decompx_ready:
|
1036 |
return pooled_output, pre_pooled_output
|
1037 |
return pooled_output
|
1038 |
|
|
|
1378 |
output_attentions: Optional[bool] = None,
|
1379 |
output_hidden_states: Optional[bool] = None,
|
1380 |
return_dict: Optional[bool] = None,
|
1381 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
1382 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
|
1383 |
r"""
|
1384 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
|
|
1477 |
output_attentions=output_attentions,
|
1478 |
output_hidden_states=output_hidden_states,
|
1479 |
return_dict=return_dict,
|
1480 |
+
decompx_config=decompx_config, # added by Fayyaz / Modarressi
|
1481 |
)
|
1482 |
sequence_output = encoder_outputs[0]
|
1483 |
+
decompx_ready = decompx_config is not None
|
1484 |
+
pooled_output = self.pooler(sequence_output, decompx_ready=decompx_ready) if self.pooler is not None else None
|
1485 |
|
1486 |
+
if decompx_ready:
|
1487 |
pre_act_pooled = pooled_output[1]
|
1488 |
pooled_output = pooled_output[0]
|
1489 |
|
1490 |
+
if decompx_config.include_classifier_w_pooler:
|
1491 |
+
decompx_idx = -2 if decompx_config.output_all_layers else -1
|
1492 |
+
aggregated_attribution_vectors = encoder_outputs[decompx_idx].aggregated[0]
|
1493 |
|
1494 |
+
encoder_outputs[decompx_idx].aggregated = output_builder(aggregated_attribution_vectors, decompx_config.output_aggregated)
|
1495 |
|
1496 |
pooler_decomposed = self.ffn_decomposer(
|
1497 |
attribution_vectors=aggregated_attribution_vectors[:, 0],
|
1498 |
pre_act_pooled=pre_act_pooled,
|
1499 |
post_act_pooled=pooled_output,
|
1500 |
+
include_biases=decompx_config.include_biases,
|
1501 |
+
bias_decomp_type="biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type,
|
1502 |
+
tanh_approx_type=decompx_config.tanh_approx_type
|
1503 |
)
|
1504 |
|
1505 |
+
encoder_outputs[decompx_idx].pooler = pooler_decomposed
|
1506 |
|
1507 |
if not return_dict:
|
1508 |
return (sequence_output, pooled_output) + encoder_outputs[1:]
|
|
|
2085 |
output_attentions: Optional[bool] = None,
|
2086 |
output_hidden_states: Optional[bool] = None,
|
2087 |
return_dict: Optional[bool] = None,
|
2088 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
2089 |
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
|
2090 |
r"""
|
2091 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
|
|
2105 |
output_attentions=output_attentions,
|
2106 |
output_hidden_states=output_hidden_states,
|
2107 |
return_dict=return_dict,
|
2108 |
+
decompx_config=decompx_config
|
2109 |
)
|
2110 |
|
2111 |
pooled_output = outputs[1]
|
|
|
2113 |
pooled_output = self.dropout(pooled_output)
|
2114 |
logits = self.classifier(pooled_output)
|
2115 |
|
2116 |
+
if decompx_config and decompx_config.include_classifier_w_pooler:
|
2117 |
+
decompx_idx = -2 if decompx_config.output_all_layers else -1
|
2118 |
+
aggregated_attribution_vectors = outputs[decompx_idx].pooler
|
2119 |
|
2120 |
+
outputs[decompx_idx].pooler = output_builder(aggregated_attribution_vectors, decompx_config.output_pooler)
|
2121 |
|
2122 |
classifier_decomposed = self.ffn_decomposer(
|
2123 |
attribution_vectors=aggregated_attribution_vectors,
|
2124 |
+
include_biases=decompx_config.include_biases,
|
2125 |
+
bias_decomp_type="biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type
|
2126 |
)
|
2127 |
|
2128 |
+
if decompx_config.include_bias_token and decompx_config.bias_decomp_type is not None:
|
2129 |
bias_token = classifier_decomposed[:,-1,:].detach().clone()
|
2130 |
classifier_decomposed = classifier_decomposed[:,:-1,:]
|
2131 |
classifier_decomposed = self.biastoken_decomposer(
|
2132 |
bias_token,
|
2133 |
classifier_decomposed,
|
2134 |
+
bias_decomp_type=decompx_config.bias_decomp_type
|
2135 |
)
|
2136 |
|
2137 |
|
2138 |
+
outputs[decompx_idx].classifier = classifier_decomposed if decompx_config.output_classifier else None
|
2139 |
|
2140 |
loss = None
|
2141 |
if labels is not None:
|
DecompX/src/modeling_roberta.py
CHANGED
@@ -24,7 +24,7 @@ from packaging import version
|
|
24 |
from torch import nn
|
25 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
26 |
|
27 |
-
from .
|
28 |
|
29 |
from transformers.activations import ACT2FN, gelu
|
30 |
from transformers.modeling_outputs import (
|
@@ -52,7 +52,6 @@ from transformers.utils import (
|
|
52 |
)
|
53 |
from transformers.models.roberta.configuration_roberta import RobertaConfig
|
54 |
|
55 |
-
|
56 |
logger = logging.get_logger(__name__)
|
57 |
|
58 |
_CHECKPOINT_FOR_DOC = "roberta-base"
|
@@ -69,6 +68,7 @@ ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
|
69 |
# See all RoBERTa models at https://huggingface.co/models?filter=roberta
|
70 |
]
|
71 |
|
|
|
72 |
def output_builder(input_vector, output_mode):
|
73 |
if output_mode is None:
|
74 |
return None
|
@@ -119,7 +119,7 @@ class RobertaEmbeddings(nn.Module):
|
|
119 |
)
|
120 |
|
121 |
def forward(
|
122 |
-
|
123 |
):
|
124 |
if position_ids is None:
|
125 |
if input_ids is not None:
|
@@ -220,16 +220,16 @@ class RobertaSelfAttention(nn.Module):
|
|
220 |
return x.permute(0, 3, 1, 2, 4)
|
221 |
|
222 |
def forward(
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
) -> Tuple[torch.Tensor]:
|
234 |
mixed_query_layer = self.query(hidden_states)
|
235 |
|
@@ -315,7 +315,7 @@ class RobertaSelfAttention(nn.Module):
|
|
315 |
|
316 |
# added by Fayyaz / Modarressi
|
317 |
# -------------------------------
|
318 |
-
if
|
319 |
outputs = (context_layer, attention_probs, value_layer, decomposed_value_layer)
|
320 |
return outputs
|
321 |
# -------------------------------
|
@@ -336,14 +336,14 @@ class RobertaSelfOutput(nn.Module):
|
|
336 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
337 |
|
338 |
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
339 |
-
|
340 |
hidden_states = self.dense(hidden_states)
|
341 |
hidden_states = self.dropout(hidden_states)
|
342 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
343 |
pre_ln_states = hidden_states + input_tensor # added by Fayyaz / Modarressi
|
344 |
post_ln_states = self.LayerNorm(pre_ln_states) # added by Fayyaz / Modarressi
|
345 |
# added by Fayyaz / Modarressi
|
346 |
-
if
|
347 |
return post_ln_states, pre_ln_states
|
348 |
else:
|
349 |
return post_ln_states
|
@@ -376,16 +376,16 @@ class RobertaAttention(nn.Module):
|
|
376 |
self.pruned_heads = self.pruned_heads.union(heads)
|
377 |
|
378 |
def forward(
|
379 |
-
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
389 |
) -> Tuple[torch.Tensor]:
|
390 |
self_outputs = self.self(
|
391 |
hidden_states,
|
@@ -396,20 +396,21 @@ class RobertaAttention(nn.Module):
|
|
396 |
encoder_attention_mask,
|
397 |
past_key_value,
|
398 |
output_attentions,
|
399 |
-
|
400 |
)
|
401 |
attention_output = self.output(
|
402 |
self_outputs[0],
|
403 |
hidden_states,
|
404 |
-
|
405 |
)
|
406 |
|
407 |
# Added by Fayyaz / Modarressi
|
408 |
# -------------------------------
|
409 |
-
if
|
410 |
_, attention_probs, value_layer, decomposed_value_layer = self_outputs
|
411 |
attention_output, pre_ln_states = attention_output
|
412 |
-
outputs = (attention_output, attention_probs,) + (
|
|
|
413 |
return outputs
|
414 |
# -------------------------------
|
415 |
|
@@ -427,10 +428,10 @@ class RobertaIntermediate(nn.Module):
|
|
427 |
else:
|
428 |
self.intermediate_act_fn = config.hidden_act
|
429 |
|
430 |
-
def forward(self, hidden_states: torch.Tensor,
|
431 |
pre_act_hidden_states = self.dense(hidden_states)
|
432 |
hidden_states = self.intermediate_act_fn(pre_act_hidden_states)
|
433 |
-
if
|
434 |
return hidden_states, pre_act_hidden_states
|
435 |
return hidden_states, None
|
436 |
|
@@ -443,7 +444,7 @@ class RobertaOutput(nn.Module):
|
|
443 |
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
444 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
445 |
|
446 |
-
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
447 |
hidden_states = self.dense(hidden_states)
|
448 |
hidden_states = self.dropout(hidden_states)
|
449 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
@@ -452,7 +453,7 @@ class RobertaOutput(nn.Module):
|
|
452 |
# -------------------------------
|
453 |
pre_ln_states = hidden_states + input_tensor
|
454 |
hidden_states = self.LayerNorm(pre_ln_states)
|
455 |
-
if
|
456 |
return hidden_states, pre_ln_states
|
457 |
return hidden_states, None
|
458 |
# -------------------------------
|
@@ -496,55 +497,56 @@ class RobertaLayer(nn.Module):
|
|
496 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
497 |
elif bias_decomp_type == "cls":
|
498 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
499 |
-
weights[
|
500 |
elif bias_decomp_type == "dot":
|
501 |
weights = torch.einsum("bskd,d->bsk", attribution_vectors, bias)
|
502 |
elif bias_decomp_type == "biastoken":
|
503 |
attrib_shape = attribution_vectors.shape
|
504 |
if attrib_shape[1] == attrib_shape[2]:
|
505 |
-
attribution_vectors = torch.concat([attribution_vectors,
|
506 |
-
|
|
|
|
|
507 |
return attribution_vectors
|
508 |
|
509 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
510 |
weighted_bias = torch.matmul(weights.unsqueeze(dim=-1), bias.unsqueeze(dim=0))
|
511 |
return attribution_vectors + weighted_bias
|
512 |
|
513 |
-
|
514 |
-
|
515 |
mean = pre_ln_states.mean(-1, keepdim=True) # (batch, seq_len, 1) m(y=Σy_j)
|
516 |
var = (pre_ln_states - mean).pow(2).mean(-1, keepdim=True).unsqueeze(dim=2) # (batch, seq_len, 1, 1) s(y)
|
517 |
|
518 |
each_mean = attribution_vectors.mean(-1, keepdim=True) # (batch, seq_len, seq_len, 1) m(y_j)
|
519 |
|
520 |
normalized_layer = torch.div(attribution_vectors - each_mean,
|
521 |
-
|
522 |
|
523 |
post_ln_layer = torch.einsum('bskd,d->bskd', normalized_layer,
|
524 |
-
|
525 |
-
|
526 |
if include_biases:
|
527 |
return self.bias_decomposer(beta, post_ln_layer, bias_decomp_type=bias_decomp_type)
|
528 |
else:
|
529 |
-
return post_ln_layer
|
530 |
-
|
531 |
|
532 |
def gelu_linear_approximation(self, intermediate_hidden_states, intermediate_output):
|
533 |
def phi(x):
|
534 |
return (1 + torch.erf(x / math.sqrt(2))) / 2.
|
535 |
-
|
536 |
def normal_pdf(x):
|
537 |
-
return torch.exp(-(x**2) / 2) / math.sqrt(2. * math.pi)
|
538 |
|
539 |
def gelu_deriv(x):
|
540 |
-
return phi(x)+x*normal_pdf(x)
|
541 |
-
|
542 |
m = gelu_deriv(intermediate_hidden_states)
|
543 |
b = intermediate_output - m * intermediate_hidden_states
|
544 |
return m, b
|
545 |
|
546 |
-
|
547 |
-
|
548 |
m, b = self.gelu_linear_approximation(intermediate_hidden_states, intermediate_output)
|
549 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
550 |
|
@@ -559,46 +561,49 @@ class RobertaLayer(nn.Module):
|
|
559 |
weights = (torch.norm(mx, dim=-1) != 0) * 1.0
|
560 |
elif bias_decomp_type == "cls":
|
561 |
weights = torch.zeros(mx.shape[:-1], device=mx.device)
|
562 |
-
weights[
|
563 |
|
564 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
565 |
weighted_bias = torch.einsum("bsl,bsk->bskl", b, weights)
|
566 |
return mx + weighted_bias
|
567 |
|
568 |
-
|
569 |
def gelu_zo_decomposition(self, attribution_vectors, intermediate_hidden_states, intermediate_output):
|
570 |
m = intermediate_output / (intermediate_hidden_states + 1e-12)
|
571 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
572 |
return mx
|
573 |
-
|
574 |
|
575 |
-
def ffn_decomposer(self, attribution_vectors, intermediate_hidden_states, intermediate_output, include_biases=True,
|
|
|
576 |
post_first_layer = torch.einsum("ld,bskd->bskl", self.intermediate.dense.weight, attribution_vectors)
|
577 |
if include_biases:
|
578 |
-
post_first_layer = self.bias_decomposer(self.intermediate.dense.bias, post_first_layer,
|
|
|
579 |
|
580 |
if approximation_type == "ReLU":
|
581 |
mask_for_gelu_approx = (intermediate_hidden_states > 0)
|
582 |
post_act_first_layer = torch.einsum("bskl, bsl->bskl", post_first_layer, mask_for_gelu_approx)
|
583 |
post_act_first_layer = post_first_layer * mask_for_gelu_approx.unsqueeze(dim=-2)
|
584 |
elif approximation_type == "GeLU_LA":
|
585 |
-
post_act_first_layer = self.gelu_decomposition(post_first_layer, intermediate_hidden_states,
|
|
|
586 |
elif approximation_type == "GeLU_ZO":
|
587 |
-
post_act_first_layer = self.gelu_zo_decomposition(post_first_layer, intermediate_hidden_states,
|
|
|
588 |
|
589 |
post_second_layer = torch.einsum("bskl, dl->bskd", post_act_first_layer, self.output.dense.weight)
|
590 |
if include_biases:
|
591 |
-
post_second_layer = self.bias_decomposer(self.output.dense.bias, post_second_layer,
|
|
|
592 |
|
593 |
return post_second_layer
|
594 |
|
595 |
-
|
596 |
-
|
597 |
if approximation_type == "ReLU":
|
598 |
theta = (intermediate_hidden_states > 0)
|
599 |
elif approximation_type == "GeLU_ZO":
|
600 |
theta = intermediate_output / (intermediate_hidden_states + 1e-12)
|
601 |
-
|
602 |
scaled_W1 = torch.einsum("bsl,ld->bsld", theta, self.intermediate.dense.weight)
|
603 |
W_equiv = torch.einsum("bsld, zl->bszd", scaled_W1, self.output.dense.weight)
|
604 |
|
@@ -625,21 +630,20 @@ class RobertaLayer(nn.Module):
|
|
625 |
post_ffn_layer = post_ffn_layer + weighted_bias
|
626 |
|
627 |
return post_ffn_layer
|
628 |
-
|
629 |
|
630 |
def forward(
|
631 |
-
|
632 |
-
|
633 |
-
|
634 |
-
|
635 |
-
|
636 |
-
|
637 |
-
|
638 |
-
|
639 |
-
|
640 |
-
|
641 |
) -> Tuple[torch.Tensor]:
|
642 |
-
|
643 |
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
644 |
# self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
645 |
# self_attention_outputs = self.attention(
|
@@ -649,7 +653,7 @@ class RobertaLayer(nn.Module):
|
|
649 |
# head_mask,
|
650 |
# output_attentions=output_attentions,
|
651 |
# past_key_value=self_attn_past_key_value,
|
652 |
-
#
|
653 |
# )
|
654 |
self_attention_outputs = self.attention(
|
655 |
hidden_states,
|
@@ -657,7 +661,7 @@ class RobertaLayer(nn.Module):
|
|
657 |
attention_mask,
|
658 |
head_mask,
|
659 |
output_attentions=output_attentions,
|
660 |
-
|
661 |
) # changed by Goro Kobayashi
|
662 |
attention_output = self_attention_outputs[0]
|
663 |
|
@@ -699,22 +703,22 @@ class RobertaLayer(nn.Module):
|
|
699 |
|
700 |
# Added by Fayyaz / Modarressi
|
701 |
# -------------------------------
|
702 |
-
bias_decomp_type = "biastoken" if
|
703 |
-
intermediate_output, pre_act_hidden_states = self.intermediate(attention_output,
|
704 |
-
layer_output, pre_ln2_states = self.output(intermediate_output, attention_output,
|
705 |
-
if
|
706 |
attention_probs, value_layer, decomposed_value_layer, pre_ln_states = outputs
|
707 |
|
708 |
headmixing_weight = self.attention.output.dense.weight.view(self.all_head_size, self.num_attention_heads,
|
709 |
-
|
710 |
|
711 |
-
if decomposed_value_layer is None or
|
712 |
transformed_layer = torch.einsum('bhsv,dhv->bhsd', value_layer, headmixing_weight) # V * W^o (z=(qk)v)
|
713 |
# Make weighted vectors αf(x) from transformed vectors (transformed_layer)
|
714 |
# and attention weights (attentions):
|
715 |
# (batch, num_heads, seq_length, seq_length, all_head_size)
|
716 |
weighted_layer = torch.einsum('bhks,bhsd->bhksd', attention_probs,
|
717 |
-
|
718 |
# Sum each weighted vectors αf(x) over all heads:
|
719 |
# (batch, seq_length, seq_length, all_head_size)
|
720 |
summed_weighted_layer = weighted_layer.sum(dim=1) # sum over heads
|
@@ -732,36 +736,38 @@ class RobertaLayer(nn.Module):
|
|
732 |
transformed_layer = torch.einsum('bhsqv,dhv->bhsqd', decomposed_value_layer, headmixing_weight)
|
733 |
|
734 |
weighted_layer = torch.einsum('bhks,bhsqd->bhkqd', attention_probs,
|
735 |
-
|
736 |
|
737 |
summed_weighted_layer = weighted_layer.sum(dim=1) # sum over heads
|
738 |
|
739 |
residual_weighted_layer = summed_weighted_layer + attribution_vectors
|
740 |
-
accumulated_bias = torch.matmul(self.attention.output.dense.weight,
|
|
|
741 |
|
742 |
-
if
|
743 |
-
residual_weighted_layer = self.bias_decomposer(accumulated_bias, residual_weighted_layer,
|
|
|
744 |
|
745 |
-
if
|
746 |
post_ln_layer = self.ln_decomposer(
|
747 |
attribution_vectors=residual_weighted_layer,
|
748 |
pre_ln_states=pre_ln_states,
|
749 |
gamma=self.attention.output.LayerNorm.weight.data,
|
750 |
beta=self.attention.output.LayerNorm.bias.data,
|
751 |
eps=self.attention.output.LayerNorm.eps,
|
752 |
-
include_biases=
|
753 |
bias_decomp_type=bias_decomp_type
|
754 |
)
|
755 |
else:
|
756 |
post_ln_layer = residual_weighted_layer
|
757 |
|
758 |
-
if
|
759 |
-
post_ffn_layer = self.ffn_decomposer_fast if
|
760 |
attribution_vectors=post_ln_layer,
|
761 |
intermediate_hidden_states=pre_act_hidden_states,
|
762 |
intermediate_output=intermediate_output,
|
763 |
-
approximation_type=
|
764 |
-
include_biases=
|
765 |
bias_decomp_type=bias_decomp_type
|
766 |
)
|
767 |
pre_ln2_layer = post_ln_layer + post_ffn_layer
|
@@ -769,25 +775,25 @@ class RobertaLayer(nn.Module):
|
|
769 |
pre_ln2_layer = post_ln_layer
|
770 |
post_ffn_layer = None
|
771 |
|
772 |
-
if
|
773 |
post_ln2_layer = self.ln_decomposer(
|
774 |
attribution_vectors=pre_ln2_layer,
|
775 |
pre_ln_states=pre_ln2_states,
|
776 |
gamma=self.output.LayerNorm.weight.data,
|
777 |
beta=self.output.LayerNorm.bias.data,
|
778 |
eps=self.output.LayerNorm.eps,
|
779 |
-
include_biases=
|
780 |
bias_decomp_type=bias_decomp_type
|
781 |
)
|
782 |
else:
|
783 |
post_ln2_layer = pre_ln2_layer
|
784 |
|
785 |
-
new_outputs =
|
786 |
-
attention=output_builder(summed_weighted_layer,
|
787 |
-
res1=output_builder(residual_weighted_layer,
|
788 |
-
LN1=output_builder(post_ln_layer,
|
789 |
-
FFN=output_builder(post_ffn_layer,
|
790 |
-
res2=output_builder(pre_ln2_layer,
|
791 |
encoder=output_builder(post_ln2_layer, "both")
|
792 |
)
|
793 |
return (layer_output,) + (new_outputs,)
|
@@ -810,18 +816,18 @@ class RobertaEncoder(nn.Module):
|
|
810 |
self.gradient_checkpointing = False
|
811 |
|
812 |
def forward(
|
813 |
-
|
814 |
-
|
815 |
-
|
816 |
-
|
817 |
-
|
818 |
-
|
819 |
-
|
820 |
-
|
821 |
-
|
822 |
-
|
823 |
-
|
824 |
-
|
825 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
|
826 |
all_hidden_states = () if output_hidden_states else None
|
827 |
all_self_attentions = () if output_attentions else None
|
@@ -829,22 +835,22 @@ class RobertaEncoder(nn.Module):
|
|
829 |
|
830 |
next_decoder_cache = () if use_cache else None
|
831 |
|
832 |
-
aggregated_encoder_norms = None
|
833 |
-
aggregated_encoder_vectors = None
|
834 |
|
835 |
# -- added by Fayyaz / Modarressi
|
836 |
-
if
|
837 |
-
|
838 |
-
attention=() if
|
839 |
-
res1=() if
|
840 |
-
LN1=() if
|
841 |
-
FFN=() if
|
842 |
-
res2=() if
|
843 |
-
encoder=() if
|
844 |
-
aggregated=() if
|
845 |
)
|
846 |
else:
|
847 |
-
|
848 |
# -- added by Fayyaz / Modarressi
|
849 |
|
850 |
for i, layer_module in enumerate(self.layer):
|
@@ -886,7 +892,7 @@ class RobertaEncoder(nn.Module):
|
|
886 |
encoder_attention_mask,
|
887 |
past_key_value,
|
888 |
output_attentions,
|
889 |
-
|
890 |
)
|
891 |
|
892 |
hidden_states = layer_outputs[0]
|
@@ -898,47 +904,52 @@ class RobertaEncoder(nn.Module):
|
|
898 |
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
|
899 |
|
900 |
# added by Fayyaz / Modarressi
|
901 |
-
if
|
902 |
-
|
903 |
-
if
|
904 |
-
if
|
905 |
raise Exception("Classifier and pooler could be included in vector aggregation mode")
|
906 |
|
907 |
-
encoder_norms =
|
908 |
|
909 |
if aggregated_encoder_norms is None:
|
910 |
-
aggregated_encoder_norms = encoder_norms * torch.exp(attention_mask).view(
|
|
|
911 |
else:
|
912 |
aggregated_encoder_norms = torch.einsum("ijk,ikm->ijm", encoder_norms, aggregated_encoder_norms)
|
913 |
-
|
914 |
-
if globenc_config.output_aggregated == "norm":
|
915 |
-
globenc_output.aggregated = (aggregated_encoder_norms,)
|
916 |
-
elif globenc_config.output_aggregated is not None:
|
917 |
-
raise Exception("Rollout aggregated values are only available in norms. Set output_aggregated to 'norm'.")
|
918 |
-
|
919 |
|
920 |
-
|
921 |
-
|
922 |
-
|
923 |
-
|
924 |
-
|
925 |
-
else:
|
926 |
-
globenc_output.aggregated = output_builder(aggregated_encoder_vectors, globenc_config.output_aggregated)
|
927 |
|
928 |
-
globenc_output.encoder = output_builder(globenc_output.encoder[0][1], globenc_config.output_encoder)
|
929 |
|
930 |
-
|
931 |
-
|
932 |
-
all_globenc_outputs.res1 = all_globenc_outputs.res1 + globenc_output.res1 if globenc_config.output_res1 else None
|
933 |
-
all_globenc_outputs.LN1 = all_globenc_outputs.LN1 + globenc_output.LN1 if globenc_config.output_LN1 else None
|
934 |
-
all_globenc_outputs.FFN = all_globenc_outputs.FFN + globenc_output.FFN if globenc_config.output_FFN else None
|
935 |
-
all_globenc_outputs.res2 = all_globenc_outputs.res2 + globenc_output.res2 if globenc_config.output_res2 else None
|
936 |
-
all_globenc_outputs.encoder = all_globenc_outputs.encoder + globenc_output.encoder if globenc_config.output_encoder else None
|
937 |
|
938 |
-
if
|
939 |
-
|
940 |
else:
|
941 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
942 |
|
943 |
if output_hidden_states:
|
944 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
@@ -952,8 +963,8 @@ class RobertaEncoder(nn.Module):
|
|
952 |
all_hidden_states,
|
953 |
all_self_attentions,
|
954 |
all_cross_attentions,
|
955 |
-
|
956 |
-
|
957 |
]
|
958 |
if v is not None
|
959 |
)
|
@@ -1147,21 +1158,21 @@ class RobertaModel(RobertaPreTrainedModel):
|
|
1147 |
)
|
1148 |
# Copied from transformers.models.bert.modeling_bert.BertModel.forward
|
1149 |
def forward(
|
1150 |
-
|
1151 |
-
|
1152 |
-
|
1153 |
-
|
1154 |
-
|
1155 |
-
|
1156 |
-
|
1157 |
-
|
1158 |
-
|
1159 |
-
|
1160 |
-
|
1161 |
-
|
1162 |
-
|
1163 |
-
|
1164 |
-
|
1165 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
|
1166 |
r"""
|
1167 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
@@ -1260,7 +1271,7 @@ class RobertaModel(RobertaPreTrainedModel):
|
|
1260 |
output_attentions=output_attentions,
|
1261 |
output_hidden_states=output_hidden_states,
|
1262 |
return_dict=return_dict,
|
1263 |
-
|
1264 |
)
|
1265 |
sequence_output = encoder_outputs[0]
|
1266 |
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
|
@@ -1310,21 +1321,21 @@ class RobertaForCausalLM(RobertaPreTrainedModel):
|
|
1310 |
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1311 |
@replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
|
1312 |
def forward(
|
1313 |
-
|
1314 |
-
|
1315 |
-
|
1316 |
-
|
1317 |
-
|
1318 |
-
|
1319 |
-
|
1320 |
-
|
1321 |
-
|
1322 |
-
|
1323 |
-
|
1324 |
-
|
1325 |
-
|
1326 |
-
|
1327 |
-
|
1328 |
) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
|
1329 |
r"""
|
1330 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
@@ -1473,19 +1484,19 @@ class RobertaForMaskedLM(RobertaPreTrainedModel):
|
|
1473 |
expected_loss=0.1,
|
1474 |
)
|
1475 |
def forward(
|
1476 |
-
|
1477 |
-
|
1478 |
-
|
1479 |
-
|
1480 |
-
|
1481 |
-
|
1482 |
-
|
1483 |
-
|
1484 |
-
|
1485 |
-
|
1486 |
-
|
1487 |
-
|
1488 |
-
|
1489 |
) -> Union[Tuple, MaskedLMOutput]:
|
1490 |
r"""
|
1491 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
@@ -1580,8 +1591,8 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1580 |
|
1581 |
def tanh_linear_approximation(self, pre_act_pooled, post_act_pooled):
|
1582 |
def tanh_deriv(x):
|
1583 |
-
return 1 - torch.tanh(x)**2.0
|
1584 |
-
|
1585 |
m = tanh_deriv(pre_act_pooled)
|
1586 |
b = post_act_pooled - m * pre_act_pooled
|
1587 |
return m, b
|
@@ -1601,7 +1612,7 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1601 |
weights = (torch.norm(mx, dim=-1) != 0) * 1.0
|
1602 |
elif bias_decomp_type == "cls":
|
1603 |
weights = torch.zeros(mx.shape[:-1], device=mx.device)
|
1604 |
-
weights[:,0] = 1.0
|
1605 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
1606 |
weighted_bias = torch.einsum("bd,bk->bkd", b, weights)
|
1607 |
return mx + weighted_bias
|
@@ -1610,14 +1621,16 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1610 |
m = post_act_pooled / (pre_act_pooled + 1e-12)
|
1611 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
1612 |
return mx
|
1613 |
-
|
1614 |
-
def pooler_decomposer(self, attribution_vectors, pre_act_pooled, post_act_pooled, include_biases=True,
|
|
|
1615 |
post_pool = torch.einsum("ld,bsd->bsl", self.classifier.dense.weight, attribution_vectors)
|
1616 |
if include_biases:
|
1617 |
post_pool = self.bias_decomposer(self.classifier.dense.bias, post_pool, bias_decomp_type=bias_decomp_type)
|
1618 |
|
1619 |
if tanh_approx_type == "LA":
|
1620 |
-
post_act_pool = self.tanh_la_decomposition(post_pool, pre_act_pooled, post_act_pooled,
|
|
|
1621 |
else:
|
1622 |
post_act_pool = self.tanh_zo_decomposition(post_pool, pre_act_pooled, post_act_pooled)
|
1623 |
|
@@ -1639,11 +1652,11 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1639 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
1640 |
elif bias_decomp_type == "cls":
|
1641 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
1642 |
-
weights[:,0] = 1.0
|
1643 |
elif bias_decomp_type == "dot":
|
1644 |
weights = torch.einsum("bkd,d->bk", attribution_vectors, bias)
|
1645 |
elif bias_decomp_type == "biastoken":
|
1646 |
-
attribution_vectors[
|
1647 |
return attribution_vectors
|
1648 |
|
1649 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
@@ -1666,7 +1679,7 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1666 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
1667 |
elif bias_decomp_type == "cls":
|
1668 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
1669 |
-
weights[:,0] = 1.0
|
1670 |
elif bias_decomp_type == "dot":
|
1671 |
weights = torch.einsum("bkd,d->bk", attribution_vectors, biastoken)
|
1672 |
|
@@ -1677,7 +1690,8 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1677 |
def ffn_decomposer(self, attribution_vectors, include_biases=True, bias_decomp_type="absdot"):
|
1678 |
post_classifier = torch.einsum("ld,bkd->bkl", self.classifier.out_proj.weight, attribution_vectors)
|
1679 |
if include_biases:
|
1680 |
-
post_classifier = self.bias_decomposer(self.classifier.out_proj.bias, post_classifier,
|
|
|
1681 |
|
1682 |
return post_classifier
|
1683 |
|
@@ -1691,18 +1705,18 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1691 |
expected_loss=0.08,
|
1692 |
)
|
1693 |
def forward(
|
1694 |
-
|
1695 |
-
|
1696 |
-
|
1697 |
-
|
1698 |
-
|
1699 |
-
|
1700 |
-
|
1701 |
-
|
1702 |
-
|
1703 |
-
|
1704 |
-
|
1705 |
-
|
1706 |
) -> Union[Tuple, SequenceClassifierOutput]:
|
1707 |
r"""
|
1708 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
@@ -1722,50 +1736,51 @@ class RobertaForSequenceClassification(RobertaPreTrainedModel):
|
|
1722 |
output_attentions=output_attentions,
|
1723 |
output_hidden_states=output_hidden_states,
|
1724 |
return_dict=return_dict,
|
1725 |
-
|
1726 |
)
|
1727 |
sequence_output = outputs[0]
|
1728 |
-
logits, mid_classifier_outputs = self.classifier(sequence_output,
|
1729 |
|
1730 |
-
if
|
1731 |
pre_act_pooled = mid_classifier_outputs[0]
|
1732 |
pooled_output = mid_classifier_outputs[1]
|
1733 |
|
1734 |
-
if
|
1735 |
-
|
1736 |
-
aggregated_attribution_vectors = outputs[
|
1737 |
|
1738 |
-
outputs[
|
|
|
1739 |
|
1740 |
pooler_decomposed = self.pooler_decomposer(
|
1741 |
-
attribution_vectors=aggregated_attribution_vectors[:, 0],
|
1742 |
-
pre_act_pooled=pre_act_pooled,
|
1743 |
-
post_act_pooled=pooled_output,
|
1744 |
-
include_biases=
|
1745 |
-
bias_decomp_type="biastoken" if
|
1746 |
-
tanh_approx_type=
|
1747 |
)
|
1748 |
|
1749 |
aggregated_attribution_vectors = pooler_decomposed
|
1750 |
|
1751 |
-
outputs[
|
1752 |
|
1753 |
classifier_decomposed = self.ffn_decomposer(
|
1754 |
-
attribution_vectors=aggregated_attribution_vectors,
|
1755 |
-
include_biases=
|
1756 |
-
bias_decomp_type="biastoken" if
|
1757 |
)
|
1758 |
-
|
1759 |
-
if
|
1760 |
-
bias_token = classifier_decomposed[
|
1761 |
-
classifier_decomposed = classifier_decomposed[
|
1762 |
classifier_decomposed = self.biastoken_decomposer(
|
1763 |
-
bias_token,
|
1764 |
-
classifier_decomposed,
|
1765 |
-
bias_decomp_type=
|
1766 |
)
|
1767 |
|
1768 |
-
outputs[
|
1769 |
|
1770 |
loss = None
|
1771 |
if labels is not None:
|
@@ -1830,17 +1845,17 @@ class RobertaForMultipleChoice(RobertaPreTrainedModel):
|
|
1830 |
config_class=_CONFIG_FOR_DOC,
|
1831 |
)
|
1832 |
def forward(
|
1833 |
-
|
1834 |
-
|
1835 |
-
|
1836 |
-
|
1837 |
-
|
1838 |
-
|
1839 |
-
|
1840 |
-
|
1841 |
-
|
1842 |
-
|
1843 |
-
|
1844 |
) -> Union[Tuple, MultipleChoiceModelOutput]:
|
1845 |
r"""
|
1846 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
@@ -1930,17 +1945,17 @@ class RobertaForTokenClassification(RobertaPreTrainedModel):
|
|
1930 |
expected_loss=0.01,
|
1931 |
)
|
1932 |
def forward(
|
1933 |
-
|
1934 |
-
|
1935 |
-
|
1936 |
-
|
1937 |
-
|
1938 |
-
|
1939 |
-
|
1940 |
-
|
1941 |
-
|
1942 |
-
|
1943 |
-
|
1944 |
) -> Union[Tuple, TokenClassifierOutput]:
|
1945 |
r"""
|
1946 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
@@ -1994,14 +2009,14 @@ class RobertaClassificationHead(nn.Module):
|
|
1994 |
self.dropout = nn.Dropout(classifier_dropout)
|
1995 |
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
1996 |
|
1997 |
-
def forward(self, features,
|
1998 |
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
1999 |
x = self.dropout(x)
|
2000 |
pre_act = self.dense(x)
|
2001 |
post_act = torch.tanh(pre_act)
|
2002 |
x = self.dropout(post_act)
|
2003 |
x = self.out_proj(x)
|
2004 |
-
if
|
2005 |
return x, (pre_act, post_act)
|
2006 |
return x, None
|
2007 |
|
@@ -2037,18 +2052,18 @@ class RobertaForQuestionAnswering(RobertaPreTrainedModel):
|
|
2037 |
expected_loss=0.86,
|
2038 |
)
|
2039 |
def forward(
|
2040 |
-
|
2041 |
-
|
2042 |
-
|
2043 |
-
|
2044 |
-
|
2045 |
-
|
2046 |
-
|
2047 |
-
|
2048 |
-
|
2049 |
-
|
2050 |
-
|
2051 |
-
|
2052 |
) -> Union[Tuple, QuestionAnsweringModelOutput]:
|
2053 |
r"""
|
2054 |
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
@@ -2124,4 +2139,4 @@ def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_l
|
|
2124 |
# The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
|
2125 |
mask = input_ids.ne(padding_idx).int()
|
2126 |
incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
|
2127 |
-
return incremental_indices.long() + padding_idx
|
|
|
24 |
from torch import nn
|
25 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
26 |
|
27 |
+
from .decompx_utils import DecompXConfig, DecompXOutput
|
28 |
|
29 |
from transformers.activations import ACT2FN, gelu
|
30 |
from transformers.modeling_outputs import (
|
|
|
52 |
)
|
53 |
from transformers.models.roberta.configuration_roberta import RobertaConfig
|
54 |
|
|
|
55 |
logger = logging.get_logger(__name__)
|
56 |
|
57 |
_CHECKPOINT_FOR_DOC = "roberta-base"
|
|
|
68 |
# See all RoBERTa models at https://huggingface.co/models?filter=roberta
|
69 |
]
|
70 |
|
71 |
+
|
72 |
def output_builder(input_vector, output_mode):
|
73 |
if output_mode is None:
|
74 |
return None
|
|
|
119 |
)
|
120 |
|
121 |
def forward(
|
122 |
+
self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
|
123 |
):
|
124 |
if position_ids is None:
|
125 |
if input_ids is not None:
|
|
|
220 |
return x.permute(0, 3, 1, 2, 4)
|
221 |
|
222 |
def forward(
|
223 |
+
self,
|
224 |
+
hidden_states: torch.Tensor,
|
225 |
+
attribution_vectors: Optional[torch.FloatTensor] = None,
|
226 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
227 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
228 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
229 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
230 |
+
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
231 |
+
output_attentions: Optional[bool] = False,
|
232 |
+
decompx_ready: Optional[bool] = None, # added by Fayyaz / Modarressi
|
233 |
) -> Tuple[torch.Tensor]:
|
234 |
mixed_query_layer = self.query(hidden_states)
|
235 |
|
|
|
315 |
|
316 |
# added by Fayyaz / Modarressi
|
317 |
# -------------------------------
|
318 |
+
if decompx_ready:
|
319 |
outputs = (context_layer, attention_probs, value_layer, decomposed_value_layer)
|
320 |
return outputs
|
321 |
# -------------------------------
|
|
|
336 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
337 |
|
338 |
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor,
|
339 |
+
decompx_ready=False): # added by Fayyaz / Modarressi
|
340 |
hidden_states = self.dense(hidden_states)
|
341 |
hidden_states = self.dropout(hidden_states)
|
342 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
343 |
pre_ln_states = hidden_states + input_tensor # added by Fayyaz / Modarressi
|
344 |
post_ln_states = self.LayerNorm(pre_ln_states) # added by Fayyaz / Modarressi
|
345 |
# added by Fayyaz / Modarressi
|
346 |
+
if decompx_ready:
|
347 |
return post_ln_states, pre_ln_states
|
348 |
else:
|
349 |
return post_ln_states
|
|
|
376 |
self.pruned_heads = self.pruned_heads.union(heads)
|
377 |
|
378 |
def forward(
|
379 |
+
self,
|
380 |
+
hidden_states: torch.Tensor,
|
381 |
+
attribution_vectors: Optional[torch.FloatTensor] = None,
|
382 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
383 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
384 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
385 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
386 |
+
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
387 |
+
output_attentions: Optional[bool] = False,
|
388 |
+
decompx_ready: Optional[bool] = None, # added by Fayyaz / Modarressi
|
389 |
) -> Tuple[torch.Tensor]:
|
390 |
self_outputs = self.self(
|
391 |
hidden_states,
|
|
|
396 |
encoder_attention_mask,
|
397 |
past_key_value,
|
398 |
output_attentions,
|
399 |
+
decompx_ready=decompx_ready, # added by Fayyaz / Modarressi
|
400 |
)
|
401 |
attention_output = self.output(
|
402 |
self_outputs[0],
|
403 |
hidden_states,
|
404 |
+
decompx_ready=decompx_ready, # added by Goro Kobayashi (Edited by Fayyaz / Modarressi)
|
405 |
)
|
406 |
|
407 |
# Added by Fayyaz / Modarressi
|
408 |
# -------------------------------
|
409 |
+
if decompx_ready:
|
410 |
_, attention_probs, value_layer, decomposed_value_layer = self_outputs
|
411 |
attention_output, pre_ln_states = attention_output
|
412 |
+
outputs = (attention_output, attention_probs,) + (
|
413 |
+
value_layer, decomposed_value_layer, pre_ln_states) # add attentions and norms if we output them
|
414 |
return outputs
|
415 |
# -------------------------------
|
416 |
|
|
|
428 |
else:
|
429 |
self.intermediate_act_fn = config.hidden_act
|
430 |
|
431 |
+
def forward(self, hidden_states: torch.Tensor, decompx_ready: Optional[bool] = False) -> torch.Tensor:
|
432 |
pre_act_hidden_states = self.dense(hidden_states)
|
433 |
hidden_states = self.intermediate_act_fn(pre_act_hidden_states)
|
434 |
+
if decompx_ready:
|
435 |
return hidden_states, pre_act_hidden_states
|
436 |
return hidden_states, None
|
437 |
|
|
|
444 |
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
445 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
446 |
|
447 |
+
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor, decompx_ready: Optional[bool] = False):
|
448 |
hidden_states = self.dense(hidden_states)
|
449 |
hidden_states = self.dropout(hidden_states)
|
450 |
# hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
|
|
453 |
# -------------------------------
|
454 |
pre_ln_states = hidden_states + input_tensor
|
455 |
hidden_states = self.LayerNorm(pre_ln_states)
|
456 |
+
if decompx_ready:
|
457 |
return hidden_states, pre_ln_states
|
458 |
return hidden_states, None
|
459 |
# -------------------------------
|
|
|
497 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
498 |
elif bias_decomp_type == "cls":
|
499 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
500 |
+
weights[:, :, 0] = 1.0
|
501 |
elif bias_decomp_type == "dot":
|
502 |
weights = torch.einsum("bskd,d->bsk", attribution_vectors, bias)
|
503 |
elif bias_decomp_type == "biastoken":
|
504 |
attrib_shape = attribution_vectors.shape
|
505 |
if attrib_shape[1] == attrib_shape[2]:
|
506 |
+
attribution_vectors = torch.concat([attribution_vectors,
|
507 |
+
torch.zeros((attrib_shape[0], attrib_shape[1], 1, attrib_shape[3]),
|
508 |
+
device=attribution_vectors.device)], dim=-2)
|
509 |
+
attribution_vectors[:, :, -1] = attribution_vectors[:, :, -1] + bias
|
510 |
return attribution_vectors
|
511 |
|
512 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
513 |
weighted_bias = torch.matmul(weights.unsqueeze(dim=-1), bias.unsqueeze(dim=0))
|
514 |
return attribution_vectors + weighted_bias
|
515 |
|
516 |
+
def ln_decomposer(self, attribution_vectors, pre_ln_states, gamma, beta, eps, include_biases=True,
|
517 |
+
bias_decomp_type="absdot"):
|
518 |
mean = pre_ln_states.mean(-1, keepdim=True) # (batch, seq_len, 1) m(y=Σy_j)
|
519 |
var = (pre_ln_states - mean).pow(2).mean(-1, keepdim=True).unsqueeze(dim=2) # (batch, seq_len, 1, 1) s(y)
|
520 |
|
521 |
each_mean = attribution_vectors.mean(-1, keepdim=True) # (batch, seq_len, seq_len, 1) m(y_j)
|
522 |
|
523 |
normalized_layer = torch.div(attribution_vectors - each_mean,
|
524 |
+
(var + eps) ** (1 / 2)) # (batch, seq_len, seq_len, all_head_size)
|
525 |
|
526 |
post_ln_layer = torch.einsum('bskd,d->bskd', normalized_layer,
|
527 |
+
gamma) # (batch, seq_len, seq_len, all_head_size)
|
528 |
+
|
529 |
if include_biases:
|
530 |
return self.bias_decomposer(beta, post_ln_layer, bias_decomp_type=bias_decomp_type)
|
531 |
else:
|
532 |
+
return post_ln_layer
|
|
|
533 |
|
534 |
def gelu_linear_approximation(self, intermediate_hidden_states, intermediate_output):
|
535 |
def phi(x):
|
536 |
return (1 + torch.erf(x / math.sqrt(2))) / 2.
|
537 |
+
|
538 |
def normal_pdf(x):
|
539 |
+
return torch.exp(-(x ** 2) / 2) / math.sqrt(2. * math.pi)
|
540 |
|
541 |
def gelu_deriv(x):
|
542 |
+
return phi(x) + x * normal_pdf(x)
|
543 |
+
|
544 |
m = gelu_deriv(intermediate_hidden_states)
|
545 |
b = intermediate_output - m * intermediate_hidden_states
|
546 |
return m, b
|
547 |
|
548 |
+
def gelu_decomposition(self, attribution_vectors, intermediate_hidden_states, intermediate_output,
|
549 |
+
bias_decomp_type):
|
550 |
m, b = self.gelu_linear_approximation(intermediate_hidden_states, intermediate_output)
|
551 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
552 |
|
|
|
561 |
weights = (torch.norm(mx, dim=-1) != 0) * 1.0
|
562 |
elif bias_decomp_type == "cls":
|
563 |
weights = torch.zeros(mx.shape[:-1], device=mx.device)
|
564 |
+
weights[:, :, 0] = 1.0
|
565 |
|
566 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
567 |
weighted_bias = torch.einsum("bsl,bsk->bskl", b, weights)
|
568 |
return mx + weighted_bias
|
569 |
|
|
|
570 |
def gelu_zo_decomposition(self, attribution_vectors, intermediate_hidden_states, intermediate_output):
|
571 |
m = intermediate_output / (intermediate_hidden_states + 1e-12)
|
572 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
573 |
return mx
|
|
|
574 |
|
575 |
+
def ffn_decomposer(self, attribution_vectors, intermediate_hidden_states, intermediate_output, include_biases=True,
|
576 |
+
approximation_type="GeLU_LA", bias_decomp_type="absdot"):
|
577 |
post_first_layer = torch.einsum("ld,bskd->bskl", self.intermediate.dense.weight, attribution_vectors)
|
578 |
if include_biases:
|
579 |
+
post_first_layer = self.bias_decomposer(self.intermediate.dense.bias, post_first_layer,
|
580 |
+
bias_decomp_type=bias_decomp_type)
|
581 |
|
582 |
if approximation_type == "ReLU":
|
583 |
mask_for_gelu_approx = (intermediate_hidden_states > 0)
|
584 |
post_act_first_layer = torch.einsum("bskl, bsl->bskl", post_first_layer, mask_for_gelu_approx)
|
585 |
post_act_first_layer = post_first_layer * mask_for_gelu_approx.unsqueeze(dim=-2)
|
586 |
elif approximation_type == "GeLU_LA":
|
587 |
+
post_act_first_layer = self.gelu_decomposition(post_first_layer, intermediate_hidden_states,
|
588 |
+
intermediate_output, bias_decomp_type=bias_decomp_type)
|
589 |
elif approximation_type == "GeLU_ZO":
|
590 |
+
post_act_first_layer = self.gelu_zo_decomposition(post_first_layer, intermediate_hidden_states,
|
591 |
+
intermediate_output)
|
592 |
|
593 |
post_second_layer = torch.einsum("bskl, dl->bskd", post_act_first_layer, self.output.dense.weight)
|
594 |
if include_biases:
|
595 |
+
post_second_layer = self.bias_decomposer(self.output.dense.bias, post_second_layer,
|
596 |
+
bias_decomp_type=bias_decomp_type)
|
597 |
|
598 |
return post_second_layer
|
599 |
|
600 |
+
def ffn_decomposer_fast(self, attribution_vectors, intermediate_hidden_states, intermediate_output,
|
601 |
+
include_biases=True, approximation_type="GeLU_LA", bias_decomp_type="absdot"):
|
602 |
if approximation_type == "ReLU":
|
603 |
theta = (intermediate_hidden_states > 0)
|
604 |
elif approximation_type == "GeLU_ZO":
|
605 |
theta = intermediate_output / (intermediate_hidden_states + 1e-12)
|
606 |
+
|
607 |
scaled_W1 = torch.einsum("bsl,ld->bsld", theta, self.intermediate.dense.weight)
|
608 |
W_equiv = torch.einsum("bsld, zl->bszd", scaled_W1, self.output.dense.weight)
|
609 |
|
|
|
630 |
post_ffn_layer = post_ffn_layer + weighted_bias
|
631 |
|
632 |
return post_ffn_layer
|
|
|
633 |
|
634 |
def forward(
|
635 |
+
self,
|
636 |
+
hidden_states: torch.Tensor,
|
637 |
+
attribution_vectors: Optional[torch.FloatTensor] = None,
|
638 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
639 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
640 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
641 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
642 |
+
past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
643 |
+
output_attentions: Optional[bool] = False,
|
644 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
645 |
) -> Tuple[torch.Tensor]:
|
646 |
+
decompx_ready = decompx_config is not None
|
647 |
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
648 |
# self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
649 |
# self_attention_outputs = self.attention(
|
|
|
653 |
# head_mask,
|
654 |
# output_attentions=output_attentions,
|
655 |
# past_key_value=self_attn_past_key_value,
|
656 |
+
# decompx_ready=decompx_ready,
|
657 |
# )
|
658 |
self_attention_outputs = self.attention(
|
659 |
hidden_states,
|
|
|
661 |
attention_mask,
|
662 |
head_mask,
|
663 |
output_attentions=output_attentions,
|
664 |
+
decompx_ready=decompx_ready,
|
665 |
) # changed by Goro Kobayashi
|
666 |
attention_output = self_attention_outputs[0]
|
667 |
|
|
|
703 |
|
704 |
# Added by Fayyaz / Modarressi
|
705 |
# -------------------------------
|
706 |
+
bias_decomp_type = "biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type
|
707 |
+
intermediate_output, pre_act_hidden_states = self.intermediate(attention_output, decompx_ready=decompx_ready)
|
708 |
+
layer_output, pre_ln2_states = self.output(intermediate_output, attention_output, decompx_ready=decompx_ready)
|
709 |
+
if decompx_ready:
|
710 |
attention_probs, value_layer, decomposed_value_layer, pre_ln_states = outputs
|
711 |
|
712 |
headmixing_weight = self.attention.output.dense.weight.view(self.all_head_size, self.num_attention_heads,
|
713 |
+
self.attention_head_size)
|
714 |
|
715 |
+
if decomposed_value_layer is None or decompx_config.aggregation != "vector":
|
716 |
transformed_layer = torch.einsum('bhsv,dhv->bhsd', value_layer, headmixing_weight) # V * W^o (z=(qk)v)
|
717 |
# Make weighted vectors αf(x) from transformed vectors (transformed_layer)
|
718 |
# and attention weights (attentions):
|
719 |
# (batch, num_heads, seq_length, seq_length, all_head_size)
|
720 |
weighted_layer = torch.einsum('bhks,bhsd->bhksd', attention_probs,
|
721 |
+
transformed_layer) # attention_probs(Q*K^t) * V * W^o
|
722 |
# Sum each weighted vectors αf(x) over all heads:
|
723 |
# (batch, seq_length, seq_length, all_head_size)
|
724 |
summed_weighted_layer = weighted_layer.sum(dim=1) # sum over heads
|
|
|
736 |
transformed_layer = torch.einsum('bhsqv,dhv->bhsqd', decomposed_value_layer, headmixing_weight)
|
737 |
|
738 |
weighted_layer = torch.einsum('bhks,bhsqd->bhkqd', attention_probs,
|
739 |
+
transformed_layer) # attention_probs(Q*K^t) * V * W^o
|
740 |
|
741 |
summed_weighted_layer = weighted_layer.sum(dim=1) # sum over heads
|
742 |
|
743 |
residual_weighted_layer = summed_weighted_layer + attribution_vectors
|
744 |
+
accumulated_bias = torch.matmul(self.attention.output.dense.weight,
|
745 |
+
self.attention.self.value.bias) + self.attention.output.dense.bias
|
746 |
|
747 |
+
if decompx_config.include_biases:
|
748 |
+
residual_weighted_layer = self.bias_decomposer(accumulated_bias, residual_weighted_layer,
|
749 |
+
bias_decomp_type)
|
750 |
|
751 |
+
if decompx_config.include_LN1:
|
752 |
post_ln_layer = self.ln_decomposer(
|
753 |
attribution_vectors=residual_weighted_layer,
|
754 |
pre_ln_states=pre_ln_states,
|
755 |
gamma=self.attention.output.LayerNorm.weight.data,
|
756 |
beta=self.attention.output.LayerNorm.bias.data,
|
757 |
eps=self.attention.output.LayerNorm.eps,
|
758 |
+
include_biases=decompx_config.include_biases,
|
759 |
bias_decomp_type=bias_decomp_type
|
760 |
)
|
761 |
else:
|
762 |
post_ln_layer = residual_weighted_layer
|
763 |
|
764 |
+
if decompx_config.include_FFN:
|
765 |
+
post_ffn_layer = self.ffn_decomposer_fast if decompx_config.FFN_fast_mode else self.ffn_decomposer(
|
766 |
attribution_vectors=post_ln_layer,
|
767 |
intermediate_hidden_states=pre_act_hidden_states,
|
768 |
intermediate_output=intermediate_output,
|
769 |
+
approximation_type=decompx_config.FFN_approx_type,
|
770 |
+
include_biases=decompx_config.include_biases,
|
771 |
bias_decomp_type=bias_decomp_type
|
772 |
)
|
773 |
pre_ln2_layer = post_ln_layer + post_ffn_layer
|
|
|
775 |
pre_ln2_layer = post_ln_layer
|
776 |
post_ffn_layer = None
|
777 |
|
778 |
+
if decompx_config.include_LN2:
|
779 |
post_ln2_layer = self.ln_decomposer(
|
780 |
attribution_vectors=pre_ln2_layer,
|
781 |
pre_ln_states=pre_ln2_states,
|
782 |
gamma=self.output.LayerNorm.weight.data,
|
783 |
beta=self.output.LayerNorm.bias.data,
|
784 |
eps=self.output.LayerNorm.eps,
|
785 |
+
include_biases=decompx_config.include_biases,
|
786 |
bias_decomp_type=bias_decomp_type
|
787 |
)
|
788 |
else:
|
789 |
post_ln2_layer = pre_ln2_layer
|
790 |
|
791 |
+
new_outputs = DecompXOutput(
|
792 |
+
attention=output_builder(summed_weighted_layer, decompx_config.output_attention),
|
793 |
+
res1=output_builder(residual_weighted_layer, decompx_config.output_res1),
|
794 |
+
LN1=output_builder(post_ln_layer, decompx_config.output_res2),
|
795 |
+
FFN=output_builder(post_ffn_layer, decompx_config.output_FFN),
|
796 |
+
res2=output_builder(pre_ln2_layer, decompx_config.output_res2),
|
797 |
encoder=output_builder(post_ln2_layer, "both")
|
798 |
)
|
799 |
return (layer_output,) + (new_outputs,)
|
|
|
816 |
self.gradient_checkpointing = False
|
817 |
|
818 |
def forward(
|
819 |
+
self,
|
820 |
+
hidden_states: torch.Tensor,
|
821 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
822 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
823 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
824 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
825 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
826 |
+
use_cache: Optional[bool] = None,
|
827 |
+
output_attentions: Optional[bool] = False,
|
828 |
+
output_hidden_states: Optional[bool] = False,
|
829 |
+
return_dict: Optional[bool] = True,
|
830 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
831 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
|
832 |
all_hidden_states = () if output_hidden_states else None
|
833 |
all_self_attentions = () if output_attentions else None
|
|
|
835 |
|
836 |
next_decoder_cache = () if use_cache else None
|
837 |
|
838 |
+
aggregated_encoder_norms = None # added by Fayyaz / Modarressi
|
839 |
+
aggregated_encoder_vectors = None # added by Fayyaz / Modarressi
|
840 |
|
841 |
# -- added by Fayyaz / Modarressi
|
842 |
+
if decompx_config and decompx_config.output_all_layers:
|
843 |
+
all_decompx_outputs = DecompXOutput(
|
844 |
+
attention=() if decompx_config.output_attention else None,
|
845 |
+
res1=() if decompx_config.output_res1 else None,
|
846 |
+
LN1=() if decompx_config.output_LN1 else None,
|
847 |
+
FFN=() if decompx_config.output_LN1 else None,
|
848 |
+
res2=() if decompx_config.output_res2 else None,
|
849 |
+
encoder=() if decompx_config.output_encoder else None,
|
850 |
+
aggregated=() if decompx_config.output_aggregated and decompx_config.aggregation else None,
|
851 |
)
|
852 |
else:
|
853 |
+
all_decompx_outputs = None
|
854 |
# -- added by Fayyaz / Modarressi
|
855 |
|
856 |
for i, layer_module in enumerate(self.layer):
|
|
|
892 |
encoder_attention_mask,
|
893 |
past_key_value,
|
894 |
output_attentions,
|
895 |
+
decompx_config # added by Fayyaz / Modarressi
|
896 |
)
|
897 |
|
898 |
hidden_states = layer_outputs[0]
|
|
|
904 |
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
|
905 |
|
906 |
# added by Fayyaz / Modarressi
|
907 |
+
if decompx_config:
|
908 |
+
decompx_output = layer_outputs[1]
|
909 |
+
if decompx_config.aggregation == "rollout":
|
910 |
+
if decompx_config.include_classifier_w_pooler:
|
911 |
raise Exception("Classifier and pooler could be included in vector aggregation mode")
|
912 |
|
913 |
+
encoder_norms = decompx_output.encoder[0][0]
|
914 |
|
915 |
if aggregated_encoder_norms is None:
|
916 |
+
aggregated_encoder_norms = encoder_norms * torch.exp(attention_mask).view(
|
917 |
+
(-1, attention_mask.shape[-1], 1))
|
918 |
else:
|
919 |
aggregated_encoder_norms = torch.einsum("ijk,ikm->ijm", encoder_norms, aggregated_encoder_norms)
|
|
|
|
|
|
|
|
|
|
|
|
|
920 |
|
921 |
+
if decompx_config.output_aggregated == "norm":
|
922 |
+
decompx_output.aggregated = (aggregated_encoder_norms,)
|
923 |
+
elif decompx_config.output_aggregated is not None:
|
924 |
+
raise Exception(
|
925 |
+
"Rollout aggregated values are only available in norms. Set output_aggregated to 'norm'.")
|
|
|
|
|
926 |
|
|
|
927 |
|
928 |
+
elif decompx_config.aggregation == "vector":
|
929 |
+
aggregated_encoder_vectors = decompx_output.encoder[0][1]
|
|
|
|
|
|
|
|
|
|
|
930 |
|
931 |
+
if decompx_config.include_classifier_w_pooler:
|
932 |
+
decompx_output.aggregated = (aggregated_encoder_vectors,)
|
933 |
else:
|
934 |
+
decompx_output.aggregated = output_builder(aggregated_encoder_vectors,
|
935 |
+
decompx_config.output_aggregated)
|
936 |
+
|
937 |
+
decompx_output.encoder = output_builder(decompx_output.encoder[0][1], decompx_config.output_encoder)
|
938 |
+
|
939 |
+
if decompx_config.output_all_layers:
|
940 |
+
all_decompx_outputs.attention = all_decompx_outputs.attention + decompx_output.attention if decompx_config.output_attention else None
|
941 |
+
all_decompx_outputs.res1 = all_decompx_outputs.res1 + decompx_output.res1 if decompx_config.output_res1 else None
|
942 |
+
all_decompx_outputs.LN1 = all_decompx_outputs.LN1 + decompx_output.LN1 if decompx_config.output_LN1 else None
|
943 |
+
all_decompx_outputs.FFN = all_decompx_outputs.FFN + decompx_output.FFN if decompx_config.output_FFN else None
|
944 |
+
all_decompx_outputs.res2 = all_decompx_outputs.res2 + decompx_output.res2 if decompx_config.output_res2 else None
|
945 |
+
all_decompx_outputs.encoder = all_decompx_outputs.encoder + decompx_output.encoder if decompx_config.output_encoder else None
|
946 |
+
|
947 |
+
if decompx_config.include_classifier_w_pooler and decompx_config.aggregation == "vector":
|
948 |
+
all_decompx_outputs.aggregated = all_decompx_outputs.aggregated + output_builder(
|
949 |
+
aggregated_encoder_vectors,
|
950 |
+
decompx_config.output_aggregated) if decompx_config.output_aggregated else None
|
951 |
+
else:
|
952 |
+
all_decompx_outputs.aggregated = all_decompx_outputs.aggregated + decompx_output.aggregated if decompx_config.output_aggregated else None
|
953 |
|
954 |
if output_hidden_states:
|
955 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
963 |
all_hidden_states,
|
964 |
all_self_attentions,
|
965 |
all_cross_attentions,
|
966 |
+
decompx_output if decompx_config else None,
|
967 |
+
all_decompx_outputs
|
968 |
]
|
969 |
if v is not None
|
970 |
)
|
|
|
1158 |
)
|
1159 |
# Copied from transformers.models.bert.modeling_bert.BertModel.forward
|
1160 |
def forward(
|
1161 |
+
self,
|
1162 |
+
input_ids: Optional[torch.Tensor] = None,
|
1163 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1164 |
+
token_type_ids: Optional[torch.Tensor] = None,
|
1165 |
+
position_ids: Optional[torch.Tensor] = None,
|
1166 |
+
head_mask: Optional[torch.Tensor] = None,
|
1167 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1168 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1169 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
1170 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1171 |
+
use_cache: Optional[bool] = None,
|
1172 |
+
output_attentions: Optional[bool] = None,
|
1173 |
+
output_hidden_states: Optional[bool] = None,
|
1174 |
+
return_dict: Optional[bool] = None,
|
1175 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
1176 |
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
|
1177 |
r"""
|
1178 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
|
|
1271 |
output_attentions=output_attentions,
|
1272 |
output_hidden_states=output_hidden_states,
|
1273 |
return_dict=return_dict,
|
1274 |
+
decompx_config=decompx_config, # added by Fayyaz / Modarressi
|
1275 |
)
|
1276 |
sequence_output = encoder_outputs[0]
|
1277 |
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
|
|
|
1321 |
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1322 |
@replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
|
1323 |
def forward(
|
1324 |
+
self,
|
1325 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1326 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1327 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1328 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1329 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1330 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1331 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
1332 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
1333 |
+
labels: Optional[torch.LongTensor] = None,
|
1334 |
+
past_key_values: Tuple[Tuple[torch.FloatTensor]] = None,
|
1335 |
+
use_cache: Optional[bool] = None,
|
1336 |
+
output_attentions: Optional[bool] = None,
|
1337 |
+
output_hidden_states: Optional[bool] = None,
|
1338 |
+
return_dict: Optional[bool] = None,
|
1339 |
) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
|
1340 |
r"""
|
1341 |
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
|
|
1484 |
expected_loss=0.1,
|
1485 |
)
|
1486 |
def forward(
|
1487 |
+
self,
|
1488 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1489 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1490 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1491 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1492 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1493 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1494 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
1495 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
1496 |
+
labels: Optional[torch.LongTensor] = None,
|
1497 |
+
output_attentions: Optional[bool] = None,
|
1498 |
+
output_hidden_states: Optional[bool] = None,
|
1499 |
+
return_dict: Optional[bool] = None,
|
1500 |
) -> Union[Tuple, MaskedLMOutput]:
|
1501 |
r"""
|
1502 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
|
|
1591 |
|
1592 |
def tanh_linear_approximation(self, pre_act_pooled, post_act_pooled):
|
1593 |
def tanh_deriv(x):
|
1594 |
+
return 1 - torch.tanh(x) ** 2.0
|
1595 |
+
|
1596 |
m = tanh_deriv(pre_act_pooled)
|
1597 |
b = post_act_pooled - m * pre_act_pooled
|
1598 |
return m, b
|
|
|
1612 |
weights = (torch.norm(mx, dim=-1) != 0) * 1.0
|
1613 |
elif bias_decomp_type == "cls":
|
1614 |
weights = torch.zeros(mx.shape[:-1], device=mx.device)
|
1615 |
+
weights[:, 0] = 1.0
|
1616 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
1617 |
weighted_bias = torch.einsum("bd,bk->bkd", b, weights)
|
1618 |
return mx + weighted_bias
|
|
|
1621 |
m = post_act_pooled / (pre_act_pooled + 1e-12)
|
1622 |
mx = attribution_vectors * m.unsqueeze(dim=-2)
|
1623 |
return mx
|
1624 |
+
|
1625 |
+
def pooler_decomposer(self, attribution_vectors, pre_act_pooled, post_act_pooled, include_biases=True,
|
1626 |
+
bias_decomp_type="absdot", tanh_approx_type="LA"):
|
1627 |
post_pool = torch.einsum("ld,bsd->bsl", self.classifier.dense.weight, attribution_vectors)
|
1628 |
if include_biases:
|
1629 |
post_pool = self.bias_decomposer(self.classifier.dense.bias, post_pool, bias_decomp_type=bias_decomp_type)
|
1630 |
|
1631 |
if tanh_approx_type == "LA":
|
1632 |
+
post_act_pool = self.tanh_la_decomposition(post_pool, pre_act_pooled, post_act_pooled,
|
1633 |
+
bias_decomp_type=bias_decomp_type)
|
1634 |
else:
|
1635 |
post_act_pool = self.tanh_zo_decomposition(post_pool, pre_act_pooled, post_act_pooled)
|
1636 |
|
|
|
1652 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
1653 |
elif bias_decomp_type == "cls":
|
1654 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
1655 |
+
weights[:, 0] = 1.0
|
1656 |
elif bias_decomp_type == "dot":
|
1657 |
weights = torch.einsum("bkd,d->bk", attribution_vectors, bias)
|
1658 |
elif bias_decomp_type == "biastoken":
|
1659 |
+
attribution_vectors[:, -1] = attribution_vectors[:, -1] + bias
|
1660 |
return attribution_vectors
|
1661 |
|
1662 |
weights = weights / (weights.sum(dim=-1, keepdim=True) + 1e-12)
|
|
|
1679 |
weights = (torch.norm(attribution_vectors, dim=-1) != 0) * 1.0
|
1680 |
elif bias_decomp_type == "cls":
|
1681 |
weights = torch.zeros(attribution_vectors.shape[:-1], device=attribution_vectors.device)
|
1682 |
+
weights[:, 0] = 1.0
|
1683 |
elif bias_decomp_type == "dot":
|
1684 |
weights = torch.einsum("bkd,d->bk", attribution_vectors, biastoken)
|
1685 |
|
|
|
1690 |
def ffn_decomposer(self, attribution_vectors, include_biases=True, bias_decomp_type="absdot"):
|
1691 |
post_classifier = torch.einsum("ld,bkd->bkl", self.classifier.out_proj.weight, attribution_vectors)
|
1692 |
if include_biases:
|
1693 |
+
post_classifier = self.bias_decomposer(self.classifier.out_proj.bias, post_classifier,
|
1694 |
+
bias_decomp_type=bias_decomp_type)
|
1695 |
|
1696 |
return post_classifier
|
1697 |
|
|
|
1705 |
expected_loss=0.08,
|
1706 |
)
|
1707 |
def forward(
|
1708 |
+
self,
|
1709 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1710 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1711 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1712 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1713 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1714 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1715 |
+
labels: Optional[torch.LongTensor] = None,
|
1716 |
+
output_attentions: Optional[bool] = None,
|
1717 |
+
output_hidden_states: Optional[bool] = None,
|
1718 |
+
return_dict: Optional[bool] = None,
|
1719 |
+
decompx_config: Optional[DecompXConfig] = None, # added by Fayyaz / Modarressi
|
1720 |
) -> Union[Tuple, SequenceClassifierOutput]:
|
1721 |
r"""
|
1722 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
|
|
1736 |
output_attentions=output_attentions,
|
1737 |
output_hidden_states=output_hidden_states,
|
1738 |
return_dict=return_dict,
|
1739 |
+
decompx_config=decompx_config
|
1740 |
)
|
1741 |
sequence_output = outputs[0]
|
1742 |
+
logits, mid_classifier_outputs = self.classifier(sequence_output, decompx_ready=decompx_config is not None)
|
1743 |
|
1744 |
+
if decompx_config is not None:
|
1745 |
pre_act_pooled = mid_classifier_outputs[0]
|
1746 |
pooled_output = mid_classifier_outputs[1]
|
1747 |
|
1748 |
+
if decompx_config.include_classifier_w_pooler:
|
1749 |
+
decompx_idx = -2 if decompx_config.output_all_layers else -1
|
1750 |
+
aggregated_attribution_vectors = outputs[decompx_idx].aggregated[0]
|
1751 |
|
1752 |
+
outputs[decompx_idx].aggregated = output_builder(aggregated_attribution_vectors,
|
1753 |
+
decompx_config.output_aggregated)
|
1754 |
|
1755 |
pooler_decomposed = self.pooler_decomposer(
|
1756 |
+
attribution_vectors=aggregated_attribution_vectors[:, 0],
|
1757 |
+
pre_act_pooled=pre_act_pooled,
|
1758 |
+
post_act_pooled=pooled_output,
|
1759 |
+
include_biases=decompx_config.include_biases,
|
1760 |
+
bias_decomp_type="biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type,
|
1761 |
+
tanh_approx_type=decompx_config.tanh_approx_type
|
1762 |
)
|
1763 |
|
1764 |
aggregated_attribution_vectors = pooler_decomposed
|
1765 |
|
1766 |
+
outputs[decompx_idx].pooler = output_builder(pooler_decomposed, decompx_config.output_pooler)
|
1767 |
|
1768 |
classifier_decomposed = self.ffn_decomposer(
|
1769 |
+
attribution_vectors=aggregated_attribution_vectors,
|
1770 |
+
include_biases=decompx_config.include_biases,
|
1771 |
+
bias_decomp_type="biastoken" if decompx_config.include_bias_token else decompx_config.bias_decomp_type
|
1772 |
)
|
1773 |
+
|
1774 |
+
if decompx_config.include_bias_token and decompx_config.bias_decomp_type is not None:
|
1775 |
+
bias_token = classifier_decomposed[:, -1, :].detach().clone()
|
1776 |
+
classifier_decomposed = classifier_decomposed[:, :-1, :]
|
1777 |
classifier_decomposed = self.biastoken_decomposer(
|
1778 |
+
bias_token,
|
1779 |
+
classifier_decomposed,
|
1780 |
+
bias_decomp_type=decompx_config.bias_decomp_type
|
1781 |
)
|
1782 |
|
1783 |
+
outputs[decompx_idx].classifier = classifier_decomposed if decompx_config.output_classifier else None
|
1784 |
|
1785 |
loss = None
|
1786 |
if labels is not None:
|
|
|
1845 |
config_class=_CONFIG_FOR_DOC,
|
1846 |
)
|
1847 |
def forward(
|
1848 |
+
self,
|
1849 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1850 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1851 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1852 |
+
labels: Optional[torch.LongTensor] = None,
|
1853 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1854 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1855 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1856 |
+
output_attentions: Optional[bool] = None,
|
1857 |
+
output_hidden_states: Optional[bool] = None,
|
1858 |
+
return_dict: Optional[bool] = None,
|
1859 |
) -> Union[Tuple, MultipleChoiceModelOutput]:
|
1860 |
r"""
|
1861 |
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
|
|
1945 |
expected_loss=0.01,
|
1946 |
)
|
1947 |
def forward(
|
1948 |
+
self,
|
1949 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1950 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1951 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1952 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1953 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1954 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1955 |
+
labels: Optional[torch.LongTensor] = None,
|
1956 |
+
output_attentions: Optional[bool] = None,
|
1957 |
+
output_hidden_states: Optional[bool] = None,
|
1958 |
+
return_dict: Optional[bool] = None,
|
1959 |
) -> Union[Tuple, TokenClassifierOutput]:
|
1960 |
r"""
|
1961 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
|
|
2009 |
self.dropout = nn.Dropout(classifier_dropout)
|
2010 |
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
2011 |
|
2012 |
+
def forward(self, features, decompx_ready=False, **kwargs):
|
2013 |
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
2014 |
x = self.dropout(x)
|
2015 |
pre_act = self.dense(x)
|
2016 |
post_act = torch.tanh(pre_act)
|
2017 |
x = self.dropout(post_act)
|
2018 |
x = self.out_proj(x)
|
2019 |
+
if decompx_ready:
|
2020 |
return x, (pre_act, post_act)
|
2021 |
return x, None
|
2022 |
|
|
|
2052 |
expected_loss=0.86,
|
2053 |
)
|
2054 |
def forward(
|
2055 |
+
self,
|
2056 |
+
input_ids: Optional[torch.LongTensor] = None,
|
2057 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
2058 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
2059 |
+
position_ids: Optional[torch.LongTensor] = None,
|
2060 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
2061 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
2062 |
+
start_positions: Optional[torch.LongTensor] = None,
|
2063 |
+
end_positions: Optional[torch.LongTensor] = None,
|
2064 |
+
output_attentions: Optional[bool] = None,
|
2065 |
+
output_hidden_states: Optional[bool] = None,
|
2066 |
+
return_dict: Optional[bool] = None,
|
2067 |
) -> Union[Tuple, QuestionAnsweringModelOutput]:
|
2068 |
r"""
|
2069 |
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
|
|
2139 |
# The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
|
2140 |
mask = input_ids.ne(padding_idx).int()
|
2141 |
incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
|
2142 |
+
return incremental_indices.long() + padding_idx
|