Roman Solomatin
commited on
fix dimenstions again
Browse files- config.json +2 -2
- listconranker.py +134 -75
config.json
CHANGED
@@ -12,8 +12,8 @@
|
|
12 |
"gradient_checkpointing": false,
|
13 |
"hidden_act": "gelu",
|
14 |
"hidden_dropout_prob": 0.1,
|
15 |
-
"hidden_size":
|
16 |
-
"
|
17 |
"id2label": {
|
18 |
"0": "LABEL_0"
|
19 |
},
|
|
|
12 |
"gradient_checkpointing": false,
|
13 |
"hidden_act": "gelu",
|
14 |
"hidden_dropout_prob": 0.1,
|
15 |
+
"hidden_size": 1024,
|
16 |
+
"list_con_hidden_size": 1792,
|
17 |
"id2label": {
|
18 |
"0": "LABEL_0"
|
19 |
},
|
listconranker.py
CHANGED
@@ -1,20 +1,20 @@
|
|
1 |
# Copyright 2024 Bytedance Ltd. and/or its affiliates
|
2 |
#
|
3 |
-
# Permission is hereby granted, free of charge, to any person obtaining a copy of this software
|
4 |
-
# and associated documentation files (the "Software"), to deal in the Software without
|
5 |
-
# restriction, including without limitation the rights to use, copy, modify, merge, publish,
|
6 |
-
# distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the
|
7 |
# Software is furnished to do so, subject to the following conditions:
|
8 |
#
|
9 |
-
# The above copyright notice and this permission notice shall be included in all copies or
|
10 |
# substantial portions of the Software.
|
11 |
#
|
12 |
-
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
13 |
-
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
14 |
-
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
15 |
-
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
|
16 |
-
# OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
|
17 |
-
# ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
|
18 |
# OTHER DEALINGS IN THE SOFTWARE.
|
19 |
|
20 |
import math
|
@@ -23,47 +23,45 @@ from torch import nn
|
|
23 |
from torch.nn import functional as F
|
24 |
import numpy as np
|
25 |
from transformers import (
|
26 |
-
AutoTokenizer,
|
27 |
-
is_torch_npu_available,
|
28 |
-
AutoModel,
|
29 |
-
PreTrainedModel,
|
30 |
PretrainedConfig,
|
31 |
AutoConfig,
|
32 |
BertModel,
|
33 |
-
BertConfig
|
34 |
)
|
35 |
from transformers.modeling_outputs import SequenceClassifierOutput
|
36 |
from typing import Union, List, Optional
|
37 |
|
38 |
|
39 |
-
class ListConRankerConfig(
|
40 |
"""Configuration class for ListConRanker model."""
|
41 |
-
|
42 |
model_type = "ListConRanker"
|
43 |
-
|
44 |
def __init__(
|
45 |
self,
|
46 |
list_transformer_layers: int = 2,
|
47 |
-
|
48 |
-
base_hidden_size: int = 1024,
|
49 |
num_labels: int = 1,
|
50 |
-
**kwargs
|
51 |
):
|
52 |
super().__init__(**kwargs)
|
53 |
self.list_transformer_layers = list_transformer_layers
|
54 |
-
self.
|
55 |
-
self.base_hidden_size = base_hidden_size
|
56 |
self.num_labels = num_labels
|
57 |
|
58 |
self.bert_config = BertConfig(**kwargs)
|
59 |
-
self.bert_config.hidden_size = self.base_hidden_size
|
60 |
self.bert_config.output_hidden_states = True
|
61 |
|
|
|
62 |
class QueryEmbedding(nn.Module):
|
63 |
def __init__(self, config) -> None:
|
64 |
super().__init__()
|
65 |
-
self.query_embedding = nn.Embedding(2, config.
|
66 |
-
self.layerNorm = nn.LayerNorm(config.
|
67 |
|
68 |
def forward(self, x, tags):
|
69 |
query_embeddings = self.query_embedding(tags)
|
@@ -71,40 +69,70 @@ class QueryEmbedding(nn.Module):
|
|
71 |
x = self.layerNorm(x)
|
72 |
return x
|
73 |
|
|
|
74 |
class ListTransformer(nn.Module):
|
75 |
def __init__(self, num_layer, config) -> None:
|
76 |
super().__init__()
|
77 |
self.config = config
|
78 |
-
self.list_transformer_layer = nn.TransformerEncoderLayer(
|
79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
self.relu = nn.ReLU()
|
81 |
self.query_embedding = QueryEmbedding(config)
|
82 |
|
83 |
-
self.linear_score3 = nn.Linear(
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
86 |
|
87 |
-
def forward(
|
88 |
-
|
89 |
-
|
90 |
batch_pair_features = pair_features.split(pair_nums)
|
91 |
|
92 |
pair_feature_query_passage_concat_list = []
|
93 |
for i in range(len(batch_pair_features)):
|
94 |
-
pair_feature_query =
|
|
|
|
|
95 |
pair_feature_passage = batch_pair_features[i][1:]
|
96 |
-
pair_feature_query_passage_concat_list.append(
|
97 |
-
|
|
|
|
|
|
|
|
|
98 |
|
99 |
-
batch_pair_features = nn.utils.rnn.pad_sequence(
|
|
|
|
|
100 |
|
101 |
-
query_embedding_tags = torch.zeros(
|
|
|
|
|
|
|
|
|
|
|
102 |
query_embedding_tags[:, 0] = 1
|
103 |
-
batch_pair_features = self.query_embedding(
|
|
|
|
|
104 |
|
105 |
mask = self.generate_attention_mask(pair_nums)
|
106 |
query_mask = self.generate_attention_mask_custom(pair_nums)
|
107 |
-
pair_list_features = self.list_transformer(
|
|
|
|
|
108 |
|
109 |
output_pair_list_features = []
|
110 |
output_query_list_features = []
|
@@ -112,20 +140,39 @@ class ListTransformer(nn.Module):
|
|
112 |
for idx, pair_num in enumerate(pair_nums):
|
113 |
output_pair_list_features.append(pair_list_features[idx, 1:pair_num, :])
|
114 |
output_query_list_features.append(pair_list_features[idx, 0, :])
|
115 |
-
pair_features_after_transformer_list.append(
|
|
|
|
|
116 |
|
117 |
pair_features_after_transformer_cat_query_list = []
|
118 |
for idx, pair_num in enumerate(pair_nums):
|
119 |
-
query_ft =
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
|
|
|
|
|
|
128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
return logits, torch.cat(pair_features_after_transformer_list, dim=0)
|
130 |
|
131 |
def generate_attention_mask(self, pair_num):
|
@@ -147,6 +194,7 @@ class ListConRankerModel(PreTrainedModel):
|
|
147 |
"""
|
148 |
ListConRanker model for sequence classification that's compatible with AutoModelForSequenceClassification.
|
149 |
"""
|
|
|
150 |
config_class = ListConRankerConfig
|
151 |
base_model_prefix = "listconranker"
|
152 |
|
@@ -155,14 +203,17 @@ class ListConRankerModel(PreTrainedModel):
|
|
155 |
self.config = config
|
156 |
self.num_labels = config.num_labels
|
157 |
self.hf_model = BertModel(config.bert_config)
|
158 |
-
|
159 |
self.sigmoid = nn.Sigmoid()
|
160 |
|
161 |
-
self.linear_in_embedding = nn.Linear(
|
|
|
|
|
162 |
self.list_transformer = ListTransformer(
|
163 |
-
config.list_transformer_layers,
|
164 |
-
config,
|
165 |
)
|
|
|
166 |
|
167 |
def forward(
|
168 |
self,
|
@@ -176,55 +227,63 @@ class ListConRankerModel(PreTrainedModel):
|
|
176 |
output_attentions: Optional[bool] = None,
|
177 |
output_hidden_states: Optional[bool] = None,
|
178 |
return_dict: Optional[bool] = None,
|
179 |
-
**kwargs
|
180 |
-
) -> Union[SequenceClassifierOutput, tuple]:
|
181 |
# Get device
|
182 |
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
183 |
self.list_transformer.device = device
|
184 |
-
|
185 |
# Forward through base model
|
186 |
if self.training:
|
187 |
pass
|
188 |
else:
|
189 |
ranker_out = self.hf_model(
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
|
|
198 |
last_hidden_state = ranker_out.last_hidden_state
|
199 |
|
200 |
pair_features = self.average_pooling(last_hidden_state, attention_mask)
|
201 |
pair_features = self.linear_in_embedding(pair_features)
|
202 |
|
203 |
-
logits, pair_features_after_list_transformer = self.list_transformer(
|
|
|
|
|
204 |
logits = self.sigmoid(logits)
|
205 |
|
206 |
return logits
|
207 |
|
208 |
def average_pooling(self, hidden_state, attention_mask):
|
209 |
-
extended_attention_mask =
|
|
|
|
|
|
|
|
|
210 |
masked_hidden_state = hidden_state * extended_attention_mask
|
211 |
sum_embeddings = torch.sum(masked_hidden_state, dim=1)
|
212 |
sum_mask = extended_attention_mask.sum(dim=1)
|
213 |
return sum_embeddings / sum_mask
|
214 |
|
215 |
@classmethod
|
216 |
-
def from_pretrained(
|
217 |
-
|
218 |
-
|
219 |
-
|
|
|
220 |
# Load custom weights
|
221 |
linear_path = f"{model_name_or_path}/linear_in_embedding.pt"
|
222 |
transformer_path = f"{model_name_or_path}/list_transformer.pt"
|
223 |
-
|
224 |
try:
|
225 |
model.linear_in_embedding.load_state_dict(torch.load(linear_path))
|
226 |
model.list_transformer.load_state_dict(torch.load(transformer_path))
|
227 |
except FileNotFoundError:
|
228 |
print(f"Warning: Could not load custom weights from {model_name_or_path}")
|
229 |
-
|
230 |
return model
|
|
|
1 |
# Copyright 2024 Bytedance Ltd. and/or its affiliates
|
2 |
#
|
3 |
+
# Permission is hereby granted, free of charge, to any person obtaining a copy of this software
|
4 |
+
# and associated documentation files (the "Software"), to deal in the Software without
|
5 |
+
# restriction, including without limitation the rights to use, copy, modify, merge, publish,
|
6 |
+
# distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the
|
7 |
# Software is furnished to do so, subject to the following conditions:
|
8 |
#
|
9 |
+
# The above copyright notice and this permission notice shall be included in all copies or
|
10 |
# substantial portions of the Software.
|
11 |
#
|
12 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
13 |
+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
14 |
+
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
15 |
+
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
|
16 |
+
# OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
|
17 |
+
# ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
|
18 |
# OTHER DEALINGS IN THE SOFTWARE.
|
19 |
|
20 |
import math
|
|
|
23 |
from torch.nn import functional as F
|
24 |
import numpy as np
|
25 |
from transformers import (
|
26 |
+
AutoTokenizer,
|
27 |
+
is_torch_npu_available,
|
28 |
+
AutoModel,
|
29 |
+
PreTrainedModel,
|
30 |
PretrainedConfig,
|
31 |
AutoConfig,
|
32 |
BertModel,
|
33 |
+
BertConfig,
|
34 |
)
|
35 |
from transformers.modeling_outputs import SequenceClassifierOutput
|
36 |
from typing import Union, List, Optional
|
37 |
|
38 |
|
39 |
+
class ListConRankerConfig(BertConfig):
|
40 |
"""Configuration class for ListConRanker model."""
|
41 |
+
|
42 |
model_type = "ListConRanker"
|
43 |
+
|
44 |
def __init__(
|
45 |
self,
|
46 |
list_transformer_layers: int = 2,
|
47 |
+
list_con_hidden_size: int = 1792,
|
|
|
48 |
num_labels: int = 1,
|
49 |
+
**kwargs,
|
50 |
):
|
51 |
super().__init__(**kwargs)
|
52 |
self.list_transformer_layers = list_transformer_layers
|
53 |
+
self.list_con_hidden_size = list_con_hidden_size
|
|
|
54 |
self.num_labels = num_labels
|
55 |
|
56 |
self.bert_config = BertConfig(**kwargs)
|
|
|
57 |
self.bert_config.output_hidden_states = True
|
58 |
|
59 |
+
|
60 |
class QueryEmbedding(nn.Module):
|
61 |
def __init__(self, config) -> None:
|
62 |
super().__init__()
|
63 |
+
self.query_embedding = nn.Embedding(2, config.list_con_hidden_size)
|
64 |
+
self.layerNorm = nn.LayerNorm(config.list_con_hidden_size)
|
65 |
|
66 |
def forward(self, x, tags):
|
67 |
query_embeddings = self.query_embedding(tags)
|
|
|
69 |
x = self.layerNorm(x)
|
70 |
return x
|
71 |
|
72 |
+
|
73 |
class ListTransformer(nn.Module):
|
74 |
def __init__(self, num_layer, config) -> None:
|
75 |
super().__init__()
|
76 |
self.config = config
|
77 |
+
self.list_transformer_layer = nn.TransformerEncoderLayer(
|
78 |
+
1792,
|
79 |
+
self.config.num_attention_heads,
|
80 |
+
batch_first=True,
|
81 |
+
activation=F.gelu,
|
82 |
+
norm_first=False,
|
83 |
+
)
|
84 |
+
self.list_transformer = nn.TransformerEncoder(
|
85 |
+
self.list_transformer_layer, num_layer
|
86 |
+
)
|
87 |
self.relu = nn.ReLU()
|
88 |
self.query_embedding = QueryEmbedding(config)
|
89 |
|
90 |
+
self.linear_score3 = nn.Linear(
|
91 |
+
config.list_con_hidden_size * 2, config.list_con_hidden_size
|
92 |
+
)
|
93 |
+
self.linear_score2 = nn.Linear(
|
94 |
+
config.list_con_hidden_size * 2, config.list_con_hidden_size
|
95 |
+
)
|
96 |
+
self.linear_score1 = nn.Linear(config.list_con_hidden_size * 2, 1)
|
97 |
|
98 |
+
def forward(
|
99 |
+
self, pair_features: torch.Tensor, pair_nums: List[int]
|
100 |
+
) -> torch.Tensor:
|
101 |
batch_pair_features = pair_features.split(pair_nums)
|
102 |
|
103 |
pair_feature_query_passage_concat_list = []
|
104 |
for i in range(len(batch_pair_features)):
|
105 |
+
pair_feature_query = (
|
106 |
+
batch_pair_features[i][0].unsqueeze(0).repeat(pair_nums[i] - 1, 1)
|
107 |
+
)
|
108 |
pair_feature_passage = batch_pair_features[i][1:]
|
109 |
+
pair_feature_query_passage_concat_list.append(
|
110 |
+
torch.cat([pair_feature_query, pair_feature_passage], dim=1)
|
111 |
+
)
|
112 |
+
pair_feature_query_passage_concat = torch.cat(
|
113 |
+
pair_feature_query_passage_concat_list, dim=0
|
114 |
+
)
|
115 |
|
116 |
+
batch_pair_features = nn.utils.rnn.pad_sequence(
|
117 |
+
batch_pair_features, batch_first=True
|
118 |
+
)
|
119 |
|
120 |
+
query_embedding_tags = torch.zeros(
|
121 |
+
batch_pair_features.size(0),
|
122 |
+
batch_pair_features.size(1),
|
123 |
+
dtype=torch.long,
|
124 |
+
device=self.device,
|
125 |
+
)
|
126 |
query_embedding_tags[:, 0] = 1
|
127 |
+
batch_pair_features = self.query_embedding(
|
128 |
+
batch_pair_features, query_embedding_tags
|
129 |
+
)
|
130 |
|
131 |
mask = self.generate_attention_mask(pair_nums)
|
132 |
query_mask = self.generate_attention_mask_custom(pair_nums)
|
133 |
+
pair_list_features = self.list_transformer(
|
134 |
+
batch_pair_features, src_key_padding_mask=mask, mask=query_mask
|
135 |
+
)
|
136 |
|
137 |
output_pair_list_features = []
|
138 |
output_query_list_features = []
|
|
|
140 |
for idx, pair_num in enumerate(pair_nums):
|
141 |
output_pair_list_features.append(pair_list_features[idx, 1:pair_num, :])
|
142 |
output_query_list_features.append(pair_list_features[idx, 0, :])
|
143 |
+
pair_features_after_transformer_list.append(
|
144 |
+
pair_list_features[idx, :pair_num, :]
|
145 |
+
)
|
146 |
|
147 |
pair_features_after_transformer_cat_query_list = []
|
148 |
for idx, pair_num in enumerate(pair_nums):
|
149 |
+
query_ft = (
|
150 |
+
output_query_list_features[idx].unsqueeze(0).repeat(pair_num - 1, 1)
|
151 |
+
)
|
152 |
+
pair_features_after_transformer_cat_query = torch.cat(
|
153 |
+
[query_ft, output_pair_list_features[idx]], dim=1
|
154 |
+
)
|
155 |
+
pair_features_after_transformer_cat_query_list.append(
|
156 |
+
pair_features_after_transformer_cat_query
|
157 |
+
)
|
158 |
+
pair_features_after_transformer_cat_query = torch.cat(
|
159 |
+
pair_features_after_transformer_cat_query_list, dim=0
|
160 |
+
)
|
161 |
|
162 |
+
pair_feature_query_passage_concat = self.relu(
|
163 |
+
self.linear_score2(pair_feature_query_passage_concat)
|
164 |
+
)
|
165 |
+
pair_features_after_transformer_cat_query = self.relu(
|
166 |
+
self.linear_score3(pair_features_after_transformer_cat_query)
|
167 |
+
)
|
168 |
+
final_ft = torch.cat(
|
169 |
+
[
|
170 |
+
pair_feature_query_passage_concat,
|
171 |
+
pair_features_after_transformer_cat_query,
|
172 |
+
],
|
173 |
+
dim=1,
|
174 |
+
)
|
175 |
+
logits = self.linear_score1(final_ft).squeeze()
|
176 |
return logits, torch.cat(pair_features_after_transformer_list, dim=0)
|
177 |
|
178 |
def generate_attention_mask(self, pair_num):
|
|
|
194 |
"""
|
195 |
ListConRanker model for sequence classification that's compatible with AutoModelForSequenceClassification.
|
196 |
"""
|
197 |
+
|
198 |
config_class = ListConRankerConfig
|
199 |
base_model_prefix = "listconranker"
|
200 |
|
|
|
203 |
self.config = config
|
204 |
self.num_labels = config.num_labels
|
205 |
self.hf_model = BertModel(config.bert_config)
|
206 |
+
|
207 |
self.sigmoid = nn.Sigmoid()
|
208 |
|
209 |
+
self.linear_in_embedding = nn.Linear(
|
210 |
+
config.hidden_size, config.list_con_hidden_size
|
211 |
+
)
|
212 |
self.list_transformer = ListTransformer(
|
213 |
+
config.list_transformer_layers,
|
214 |
+
config,
|
215 |
)
|
216 |
+
self.sep_token_id = 102 # [SEP] token ID
|
217 |
|
218 |
def forward(
|
219 |
self,
|
|
|
227 |
output_attentions: Optional[bool] = None,
|
228 |
output_hidden_states: Optional[bool] = None,
|
229 |
return_dict: Optional[bool] = None,
|
230 |
+
**kwargs,
|
231 |
+
) -> Union[SequenceClassifierOutput, tuple]:
|
232 |
# Get device
|
233 |
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
234 |
self.list_transformer.device = device
|
235 |
+
|
236 |
# Forward through base model
|
237 |
if self.training:
|
238 |
pass
|
239 |
else:
|
240 |
ranker_out = self.hf_model(
|
241 |
+
input_ids=input_ids,
|
242 |
+
attention_mask=attention_mask,
|
243 |
+
token_type_ids=token_type_ids,
|
244 |
+
position_ids=position_ids,
|
245 |
+
head_mask=head_mask,
|
246 |
+
inputs_embeds=inputs_embeds,
|
247 |
+
output_attentions=output_attentions,
|
248 |
+
return_dict=True,
|
249 |
+
)
|
250 |
last_hidden_state = ranker_out.last_hidden_state
|
251 |
|
252 |
pair_features = self.average_pooling(last_hidden_state, attention_mask)
|
253 |
pair_features = self.linear_in_embedding(pair_features)
|
254 |
|
255 |
+
logits, pair_features_after_list_transformer = self.list_transformer(
|
256 |
+
pair_features
|
257 |
+
)
|
258 |
logits = self.sigmoid(logits)
|
259 |
|
260 |
return logits
|
261 |
|
262 |
def average_pooling(self, hidden_state, attention_mask):
|
263 |
+
extended_attention_mask = (
|
264 |
+
attention_mask.unsqueeze(-1)
|
265 |
+
.expand(hidden_state.size())
|
266 |
+
.to(dtype=hidden_state.dtype)
|
267 |
+
)
|
268 |
masked_hidden_state = hidden_state * extended_attention_mask
|
269 |
sum_embeddings = torch.sum(masked_hidden_state, dim=1)
|
270 |
sum_mask = extended_attention_mask.sum(dim=1)
|
271 |
return sum_embeddings / sum_mask
|
272 |
|
273 |
@classmethod
|
274 |
+
def from_pretrained(
|
275 |
+
cls, model_name_or_path, config: Optional[ListConRankerConfig] = None, **kwargs
|
276 |
+
):
|
277 |
+
model = super().from_pretrained(model_name_or_path, config=config, **kwargs)
|
278 |
+
|
279 |
# Load custom weights
|
280 |
linear_path = f"{model_name_or_path}/linear_in_embedding.pt"
|
281 |
transformer_path = f"{model_name_or_path}/list_transformer.pt"
|
282 |
+
|
283 |
try:
|
284 |
model.linear_in_embedding.load_state_dict(torch.load(linear_path))
|
285 |
model.list_transformer.load_state_dict(torch.load(transformer_path))
|
286 |
except FileNotFoundError:
|
287 |
print(f"Warning: Could not load custom weights from {model_name_or_path}")
|
288 |
+
|
289 |
return model
|