OpenNLPLab
commited on
Commit
·
ec5dbd2
1
Parent(s):
a5b83ec
Upload modeling_transnormer.py
Browse files- modeling_transnormer.py +4 -136
modeling_transnormer.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
# coding=utf-8
|
2 |
# Copyright 2023 OpenNLPLab
|
3 |
#
|
4 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
@@ -12,8 +11,7 @@
|
|
12 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
# See the License for the specific language governing permissions and
|
14 |
# limitations under the License.
|
15 |
-
|
16 |
-
# coding=utf-8
|
17 |
""" PyTorch Transnormer model."""
|
18 |
import math
|
19 |
import os
|
@@ -30,7 +28,6 @@ from transformers.activations import ACT2FN
|
|
30 |
from transformers.modeling_outputs import (
|
31 |
BaseModelOutputWithPast,
|
32 |
CausalLMOutputWithPast,
|
33 |
-
SequenceClassifierOutputWithPast,
|
34 |
)
|
35 |
from transformers.modeling_utils import PreTrainedModel
|
36 |
from transformers.utils import (
|
@@ -752,6 +749,9 @@ class TransnormerModel(TransnormerPreTrainedModel):
|
|
752 |
|
753 |
if output_attentions:
|
754 |
all_self_attns += (layer_outputs[1],)
|
|
|
|
|
|
|
755 |
|
756 |
hidden_states = self.final_norm(hidden_states)
|
757 |
|
@@ -939,135 +939,3 @@ class TransnormerForCausalLM(TransnormerPreTrainedModel):
|
|
939 |
)
|
940 |
return reordered_past
|
941 |
|
942 |
-
|
943 |
-
@add_start_docstrings(
|
944 |
-
"""
|
945 |
-
The LLaMa Model transformer with a sequence classification head on top (linear layer).
|
946 |
-
|
947 |
-
[`TransnormerForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
948 |
-
(e.g. GPT-2) do.
|
949 |
-
|
950 |
-
Since it does classification on the last token, it requires to know the position of the last token. If a
|
951 |
-
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
952 |
-
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
953 |
-
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
954 |
-
each row of the batch).
|
955 |
-
""",
|
956 |
-
TRANSNORMER_START_DOCSTRING,
|
957 |
-
)
|
958 |
-
class TransnormerForSequenceClassification(TransnormerPreTrainedModel):
|
959 |
-
_keys_to_ignore_on_load_missing = [r"lm_head.weight"]
|
960 |
-
|
961 |
-
def __init__(self, config):
|
962 |
-
super().__init__(config)
|
963 |
-
self.num_labels = config.num_labels
|
964 |
-
self.model = TransnormerModel(config)
|
965 |
-
self.score = nn.Linear(config.decoder_embed_dim, self.num_labels, bias=False)
|
966 |
-
|
967 |
-
# Initialize weights and apply final processing
|
968 |
-
self.post_init()
|
969 |
-
|
970 |
-
def get_input_embeddings(self):
|
971 |
-
return self.model.embed_tokens
|
972 |
-
|
973 |
-
def set_input_embeddings(self, value):
|
974 |
-
self.model.embed_tokens = value
|
975 |
-
|
976 |
-
@add_start_docstrings_to_model_forward(TRANSNORMER_INPUTS_DOCSTRING)
|
977 |
-
def forward(
|
978 |
-
self,
|
979 |
-
input_ids: torch.LongTensor = None,
|
980 |
-
attn_mask: Optional[torch.Tensor] = None,
|
981 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
982 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
983 |
-
labels: Optional[torch.LongTensor] = None,
|
984 |
-
use_cache: Optional[bool] = None,
|
985 |
-
output_attentions: Optional[bool] = None,
|
986 |
-
output_hidden_states: Optional[bool] = None,
|
987 |
-
return_dict: Optional[bool] = None,
|
988 |
-
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
989 |
-
r"""
|
990 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
991 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
992 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
993 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
994 |
-
"""
|
995 |
-
return_dict = (
|
996 |
-
return_dict if return_dict is not None else self.config.use_return_dict
|
997 |
-
)
|
998 |
-
|
999 |
-
transformer_outputs = self.model(
|
1000 |
-
input_ids,
|
1001 |
-
attn_padding_mask=attn_mask,
|
1002 |
-
past_key_values=past_key_values,
|
1003 |
-
inputs_embeds=inputs_embeds,
|
1004 |
-
use_cache=use_cache,
|
1005 |
-
output_attentions=output_attentions,
|
1006 |
-
output_hidden_states=output_hidden_states,
|
1007 |
-
return_dict=return_dict,
|
1008 |
-
)
|
1009 |
-
hidden_states = transformer_outputs[0]
|
1010 |
-
|
1011 |
-
logits = self.score(hidden_states)
|
1012 |
-
|
1013 |
-
if input_ids is not None:
|
1014 |
-
batch_size = input_ids.shape[0]
|
1015 |
-
else:
|
1016 |
-
batch_size = inputs_embeds.shape[0]
|
1017 |
-
|
1018 |
-
if self.config.pad_token_id is None and batch_size != 1:
|
1019 |
-
raise ValueError(
|
1020 |
-
"Cannot handle batch sizes > 1 if no padding token is defined."
|
1021 |
-
)
|
1022 |
-
if self.config.pad_token_id is None:
|
1023 |
-
sequence_lengths = -1
|
1024 |
-
else:
|
1025 |
-
if input_ids is not None:
|
1026 |
-
sequence_lengths = (
|
1027 |
-
torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
|
1028 |
-
).to(logits.device)
|
1029 |
-
else:
|
1030 |
-
sequence_lengths = -1
|
1031 |
-
|
1032 |
-
pooled_logits = logits[
|
1033 |
-
torch.arange(batch_size, device=logits.device), sequence_lengths
|
1034 |
-
]
|
1035 |
-
|
1036 |
-
loss = None
|
1037 |
-
if labels is not None:
|
1038 |
-
labels = labels.to(logits.device)
|
1039 |
-
if self.config.problem_type is None:
|
1040 |
-
if self.num_labels == 1:
|
1041 |
-
self.config.problem_type = "regression"
|
1042 |
-
elif self.num_labels > 1 and (
|
1043 |
-
labels.dtype == torch.long or labels.dtype == torch.int
|
1044 |
-
):
|
1045 |
-
self.config.problem_type = "single_label_classification"
|
1046 |
-
else:
|
1047 |
-
self.config.problem_type = "multi_label_classification"
|
1048 |
-
|
1049 |
-
if self.config.problem_type == "regression":
|
1050 |
-
loss_fct = MSELoss()
|
1051 |
-
if self.num_labels == 1:
|
1052 |
-
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
1053 |
-
else:
|
1054 |
-
loss = loss_fct(pooled_logits, labels)
|
1055 |
-
elif self.config.problem_type == "single_label_classification":
|
1056 |
-
loss_fct = CrossEntropyLoss()
|
1057 |
-
loss = loss_fct(
|
1058 |
-
pooled_logits.view(-1, self.num_labels), labels.view(-1)
|
1059 |
-
)
|
1060 |
-
elif self.config.problem_type == "multi_label_classification":
|
1061 |
-
loss_fct = BCEWithLogitsLoss()
|
1062 |
-
loss = loss_fct(pooled_logits, labels)
|
1063 |
-
if not return_dict:
|
1064 |
-
output = (pooled_logits,) + transformer_outputs[1:]
|
1065 |
-
return ((loss,) + output) if loss is not None else output
|
1066 |
-
|
1067 |
-
return SequenceClassifierOutputWithPast(
|
1068 |
-
loss=loss,
|
1069 |
-
logits=pooled_logits,
|
1070 |
-
past_key_values=transformer_outputs.past_key_values,
|
1071 |
-
hidden_states=transformer_outputs.hidden_states,
|
1072 |
-
attentions=transformer_outputs.attentions,
|
1073 |
-
)
|
|
|
|
|
1 |
# Copyright 2023 OpenNLPLab
|
2 |
#
|
3 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
11 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
# See the License for the specific language governing permissions and
|
13 |
# limitations under the License.
|
14 |
+
# coding=utf-8
|
|
|
15 |
""" PyTorch Transnormer model."""
|
16 |
import math
|
17 |
import os
|
|
|
28 |
from transformers.modeling_outputs import (
|
29 |
BaseModelOutputWithPast,
|
30 |
CausalLMOutputWithPast,
|
|
|
31 |
)
|
32 |
from transformers.modeling_utils import PreTrainedModel
|
33 |
from transformers.utils import (
|
|
|
749 |
|
750 |
if output_attentions:
|
751 |
all_self_attns += (layer_outputs[1],)
|
752 |
+
|
753 |
+
# if idx == 0:
|
754 |
+
# break
|
755 |
|
756 |
hidden_states = self.final_norm(hidden_states)
|
757 |
|
|
|
939 |
)
|
940 |
return reordered_past
|
941 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|