Roman Solomatin
commited on
finish integration
Browse files- README.md +43 -7
- config.json +3 -1
- listconranker.py +201 -41
- tokenizer_config.json +1 -1
README.md
CHANGED
|
@@ -1,6 +1,9 @@
|
|
| 1 |
---
|
| 2 |
tags:
|
| 3 |
- mteb
|
|
|
|
|
|
|
|
|
|
| 4 |
model-index:
|
| 5 |
- name: ListConRanker
|
| 6 |
results:
|
|
@@ -103,9 +106,9 @@ To reduce the discrepancy between training and inference, we propose iterative i
|
|
| 103 |
|
| 104 |
## How to use
|
| 105 |
```python
|
| 106 |
-
from
|
| 107 |
|
| 108 |
-
reranker =
|
| 109 |
|
| 110 |
# [query, passages_1, passage_2, ..., passage_n]
|
| 111 |
batch = [
|
|
@@ -125,15 +128,48 @@ batch = [
|
|
| 125 |
]
|
| 126 |
|
| 127 |
# for conventional inference, please manage the batch size by yourself
|
| 128 |
-
scores = reranker.
|
| 129 |
print(scores)
|
| 130 |
# [[0.5126953125, 0.331298828125, 0.3642578125], [0.63671875, 0.71630859375, 0.42822265625, 0.35302734375]]
|
| 131 |
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
print(scores)
|
| 136 |
-
# [0.
|
| 137 |
```
|
| 138 |
|
| 139 |
To reproduce the results with iterative inference, please run:
|
|
|
|
| 1 |
---
|
| 2 |
tags:
|
| 3 |
- mteb
|
| 4 |
+
- sentence-transformers
|
| 5 |
+
- transformers
|
| 6 |
+
pipeline_tag: text-ranking
|
| 7 |
model-index:
|
| 8 |
- name: ListConRanker
|
| 9 |
results:
|
|
|
|
| 106 |
|
| 107 |
## How to use
|
| 108 |
```python
|
| 109 |
+
from transfoermers import AutoModelForSequenceClassification
|
| 110 |
|
| 111 |
+
reranker = AutoModelForSequenceClassification('ByteDance/ListConRanker', trust_remote_code=True)
|
| 112 |
|
| 113 |
# [query, passages_1, passage_2, ..., passage_n]
|
| 114 |
batch = [
|
|
|
|
| 128 |
]
|
| 129 |
|
| 130 |
# for conventional inference, please manage the batch size by yourself
|
| 131 |
+
scores = reranker.multi_passage(batch)
|
| 132 |
print(scores)
|
| 133 |
# [[0.5126953125, 0.331298828125, 0.3642578125], [0.63671875, 0.71630859375, 0.42822265625, 0.35302734375]]
|
| 134 |
|
| 135 |
+
inputs = tokenizer(
|
| 136 |
+
[
|
| 137 |
+
[
|
| 138 |
+
"query 1",
|
| 139 |
+
"passage_11",
|
| 140 |
+
],
|
| 141 |
+
[
|
| 142 |
+
"query_2",
|
| 143 |
+
"passage_21",
|
| 144 |
+
]
|
| 145 |
+
|
| 146 |
+
],
|
| 147 |
+
return_tensors="pt",
|
| 148 |
+
padding=True,
|
| 149 |
+
)
|
| 150 |
+
probs, logits = model(**inputs)
|
| 151 |
+
print(probs)
|
| 152 |
+
# tensor([[0.4359], [0.3840]], grad_fn=<ViewBackward0>)
|
| 153 |
+
```
|
| 154 |
+
or using the `sentence_transformers` library:
|
| 155 |
+
```python
|
| 156 |
+
from sentence_transformers import CrossEncoder
|
| 157 |
+
|
| 158 |
+
model = CrossEncoder('ByteDance/ListConRanker', trust_remote_code=True)
|
| 159 |
+
|
| 160 |
+
inputs = [
|
| 161 |
+
[
|
| 162 |
+
"query 1",
|
| 163 |
+
"passage_11",
|
| 164 |
+
],
|
| 165 |
+
[
|
| 166 |
+
"query_2",
|
| 167 |
+
"passage_21",
|
| 168 |
+
]
|
| 169 |
+
]
|
| 170 |
+
scores = model.predict(inputs)
|
| 171 |
print(scores)
|
| 172 |
+
# [0.4359, 0.3840, 0.3231]
|
| 173 |
```
|
| 174 |
|
| 175 |
To reproduce the results with iterative inference, please run:
|
config.json
CHANGED
|
@@ -39,5 +39,7 @@
|
|
| 39 |
"transformers_version": "4.45.2",
|
| 40 |
"type_vocab_size": 2,
|
| 41 |
"use_cache": true,
|
| 42 |
-
"vocab_size": 21128
|
|
|
|
|
|
|
| 43 |
}
|
|
|
|
| 39 |
"transformers_version": "4.45.2",
|
| 40 |
"type_vocab_size": 2,
|
| 41 |
"use_cache": true,
|
| 42 |
+
"vocab_size": 21128,
|
| 43 |
+
"cls_token_id": 101,
|
| 44 |
+
"sep_token_id": 102
|
| 45 |
}
|
listconranker.py
CHANGED
|
@@ -16,22 +16,18 @@
|
|
| 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
|
| 21 |
import torch
|
| 22 |
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 |
|
|
@@ -46,12 +42,16 @@ class ListConRankerConfig(BertConfig):
|
|
| 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
|
|
@@ -75,7 +75,7 @@ class ListTransformer(nn.Module):
|
|
| 75 |
super().__init__()
|
| 76 |
self.config = config
|
| 77 |
self.list_transformer_layer = nn.TransformerEncoderLayer(
|
| 78 |
-
|
| 79 |
self.config.num_attention_heads,
|
| 80 |
batch_first=True,
|
| 81 |
activation=F.gelu,
|
|
@@ -213,11 +213,10 @@ class ListConRankerModel(PreTrainedModel):
|
|
| 213 |
config.list_transformer_layers,
|
| 214 |
config,
|
| 215 |
)
|
| 216 |
-
self.sep_token_id = 102 # [SEP] token ID
|
| 217 |
|
| 218 |
def forward(
|
| 219 |
self,
|
| 220 |
-
input_ids:
|
| 221 |
attention_mask: Optional[torch.Tensor] = None,
|
| 222 |
token_type_ids: Optional[torch.Tensor] = None,
|
| 223 |
position_ids: Optional[torch.Tensor] = None,
|
|
@@ -228,36 +227,157 @@ class ListConRankerModel(PreTrainedModel):
|
|
| 228 |
output_hidden_states: Optional[bool] = None,
|
| 229 |
return_dict: Optional[bool] = None,
|
| 230 |
**kwargs,
|
| 231 |
-
) -> Union[
|
| 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 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 254 |
|
| 255 |
-
|
| 256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 257 |
)
|
| 258 |
-
logits = self.sigmoid(logits)
|
| 259 |
|
| 260 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 261 |
|
| 262 |
def average_pooling(self, hidden_state, attention_mask):
|
| 263 |
extended_attention_mask = (
|
|
@@ -275,15 +395,55 @@ class ListConRankerModel(PreTrainedModel):
|
|
| 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 |
-
|
| 280 |
-
|
| 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 |
-
|
| 288 |
|
| 289 |
return model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 |
+
from __future__ import annotations
|
| 20 |
|
|
|
|
| 21 |
import torch
|
| 22 |
from torch import nn
|
| 23 |
from torch.nn import functional as F
|
|
|
|
| 24 |
from transformers import (
|
|
|
|
|
|
|
|
|
|
| 25 |
PreTrainedModel,
|
|
|
|
|
|
|
| 26 |
BertModel,
|
| 27 |
BertConfig,
|
| 28 |
+
AutoTokenizer,
|
| 29 |
)
|
| 30 |
+
import os
|
| 31 |
from transformers.modeling_outputs import SequenceClassifierOutput
|
| 32 |
from typing import Union, List, Optional
|
| 33 |
|
|
|
|
| 42 |
list_transformer_layers: int = 2,
|
| 43 |
list_con_hidden_size: int = 1792,
|
| 44 |
num_labels: int = 1,
|
| 45 |
+
cls_token_id: int = 101,
|
| 46 |
+
sep_token_id: int = 102,
|
| 47 |
**kwargs,
|
| 48 |
):
|
| 49 |
super().__init__(**kwargs)
|
| 50 |
self.list_transformer_layers = list_transformer_layers
|
| 51 |
self.list_con_hidden_size = list_con_hidden_size
|
| 52 |
self.num_labels = num_labels
|
| 53 |
+
self.cls_token_id = cls_token_id
|
| 54 |
+
self.sep_token_id = sep_token_id
|
| 55 |
|
| 56 |
self.bert_config = BertConfig(**kwargs)
|
| 57 |
self.bert_config.output_hidden_states = True
|
|
|
|
| 75 |
super().__init__()
|
| 76 |
self.config = config
|
| 77 |
self.list_transformer_layer = nn.TransformerEncoderLayer(
|
| 78 |
+
config.list_con_hidden_size,
|
| 79 |
self.config.num_attention_heads,
|
| 80 |
batch_first=True,
|
| 81 |
activation=F.gelu,
|
|
|
|
| 213 |
config.list_transformer_layers,
|
| 214 |
config,
|
| 215 |
)
|
|
|
|
| 216 |
|
| 217 |
def forward(
|
| 218 |
self,
|
| 219 |
+
input_ids: torch.Tensor,
|
| 220 |
attention_mask: Optional[torch.Tensor] = None,
|
| 221 |
token_type_ids: Optional[torch.Tensor] = None,
|
| 222 |
position_ids: Optional[torch.Tensor] = None,
|
|
|
|
| 227 |
output_hidden_states: Optional[bool] = None,
|
| 228 |
return_dict: Optional[bool] = None,
|
| 229 |
**kwargs,
|
| 230 |
+
) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 231 |
if self.training:
|
| 232 |
+
raise NotImplementedError("Training not supported; use eval mode.")
|
| 233 |
+
device = input_ids.device
|
| 234 |
+
self.list_transformer.device = device
|
| 235 |
+
# Reorganize by unique queries and their passages
|
| 236 |
+
(
|
| 237 |
+
reorganized_input_ids,
|
| 238 |
+
reorganized_attention_mask,
|
| 239 |
+
reorganized_token_type_ids,
|
| 240 |
+
pair_nums,
|
| 241 |
+
group_indices,
|
| 242 |
+
) = self._reorganize_inputs(input_ids, attention_mask, token_type_ids)
|
| 243 |
+
|
| 244 |
+
out = self.hf_model(
|
| 245 |
+
input_ids=reorganized_input_ids,
|
| 246 |
+
attention_mask=reorganized_attention_mask,
|
| 247 |
+
token_type_ids=reorganized_token_type_ids,
|
| 248 |
+
return_dict=True,
|
| 249 |
+
)
|
| 250 |
+
feats = out.last_hidden_state
|
| 251 |
+
pooled = self.average_pooling(feats, reorganized_attention_mask)
|
| 252 |
+
embedded = self.linear_in_embedding(pooled)
|
| 253 |
+
logits, _ = self.list_transformer(embedded, pair_nums)
|
| 254 |
+
probs = self.sigmoid(logits)
|
| 255 |
+
|
| 256 |
+
# Restore original order
|
| 257 |
+
sorted_probs = self._restore_original_order(probs, group_indices)
|
| 258 |
+
sorted_logits = self._restore_original_order(logits, group_indices)
|
| 259 |
+
if not return_dict:
|
| 260 |
+
return (sorted_probs, sorted_logits)
|
| 261 |
+
|
| 262 |
+
return SequenceClassifierOutput(
|
| 263 |
+
loss=None,
|
| 264 |
+
logits=sorted_logits,
|
| 265 |
+
hidden_states=out.hidden_states,
|
| 266 |
+
attentions=out.attentions,
|
| 267 |
+
)
|
| 268 |
|
| 269 |
+
def _reorganize_inputs(
|
| 270 |
+
self,
|
| 271 |
+
input_ids: torch.Tensor,
|
| 272 |
+
attention_mask: torch.Tensor,
|
| 273 |
+
token_type_ids: Optional[torch.Tensor],
|
| 274 |
+
) -> tuple[
|
| 275 |
+
torch.Tensor, torch.Tensor, Optional[torch.Tensor], List[int], List[List[int]]
|
| 276 |
+
]:
|
| 277 |
+
"""
|
| 278 |
+
Group inputs by unique queries: for each query, produce [query] + its passages,
|
| 279 |
+
then flatten, pad, and return pair sizes and original indices mapping.
|
| 280 |
+
"""
|
| 281 |
+
batch_size = input_ids.size(0)
|
| 282 |
+
# Structure: query_key -> {
|
| 283 |
+
# 'query': (seq, mask, tt),
|
| 284 |
+
# 'passages': [(seq, mask, tt), ...],
|
| 285 |
+
# 'indices': [original_index, ...]
|
| 286 |
+
# }
|
| 287 |
+
grouped = {}
|
| 288 |
+
|
| 289 |
+
for idx in range(batch_size):
|
| 290 |
+
seq = input_ids[idx]
|
| 291 |
+
mask = attention_mask[idx]
|
| 292 |
+
token_type_ids[idx] if token_type_ids is not None else torch.zeros_like(seq)
|
| 293 |
+
|
| 294 |
+
sep_idxs = (seq == self.config.sep_token_id).nonzero(as_tuple=True)[0]
|
| 295 |
+
if sep_idxs.numel() == 0:
|
| 296 |
+
raise ValueError(f"No SEP in sequence {idx}")
|
| 297 |
+
first_sep = sep_idxs[0].item()
|
| 298 |
+
|
| 299 |
+
# Extract query and passage
|
| 300 |
+
q_seq = seq[: first_sep + 1]
|
| 301 |
+
q_mask = mask[: first_sep + 1]
|
| 302 |
+
q_tt = torch.zeros_like(q_seq)
|
| 303 |
+
|
| 304 |
+
p_seq = seq[first_sep:]
|
| 305 |
+
p_mask = mask[first_sep:]
|
| 306 |
+
p_seq = p_seq.clone()
|
| 307 |
+
p_seq[0] = self.config.cls_token_id
|
| 308 |
+
p_tt = torch.zeros_like(p_seq)
|
| 309 |
+
|
| 310 |
+
# Build key excluding CLS/SEP
|
| 311 |
+
key = tuple(
|
| 312 |
+
q_seq[
|
| 313 |
+
(q_seq != self.config.cls_token_id)
|
| 314 |
+
& (q_seq != self.config.sep_token_id)
|
| 315 |
+
].tolist()
|
| 316 |
)
|
|
|
|
| 317 |
|
| 318 |
+
if key not in grouped:
|
| 319 |
+
grouped[key] = {
|
| 320 |
+
"query": (q_seq, q_mask, q_tt),
|
| 321 |
+
"passages": [],
|
| 322 |
+
"indices": [],
|
| 323 |
+
}
|
| 324 |
+
grouped[key]["passages"].append((p_seq, p_mask, p_tt))
|
| 325 |
+
grouped[key]["indices"].append(idx)
|
| 326 |
+
|
| 327 |
+
# Flatten according to group insertion order
|
| 328 |
+
seqs, masks, tts, pair_nums, group_indices = [], [], [], [], []
|
| 329 |
+
for key, data in grouped.items():
|
| 330 |
+
q_seq, q_mask, q_tt = data["query"]
|
| 331 |
+
passages = data["passages"]
|
| 332 |
+
indices = data["indices"]
|
| 333 |
+
# record sizes and original positions
|
| 334 |
+
pair_nums.append(len(passages) + 1) # +1 for the query
|
| 335 |
+
group_indices.append(indices)
|
| 336 |
+
|
| 337 |
+
# append query then its passages
|
| 338 |
+
seqs.append(q_seq)
|
| 339 |
+
masks.append(q_mask)
|
| 340 |
+
tts.append(q_tt)
|
| 341 |
+
for p_seq, p_mask, p_tt in passages:
|
| 342 |
+
seqs.append(p_seq)
|
| 343 |
+
masks.append(p_mask)
|
| 344 |
+
tts.append(p_tt)
|
| 345 |
+
|
| 346 |
+
# Pad to uniform length
|
| 347 |
+
max_len = max(s.size(0) for s in seqs)
|
| 348 |
+
padded_seqs, padded_masks, padded_tts = [], [], []
|
| 349 |
+
for s, m, t in zip(seqs, masks, tts):
|
| 350 |
+
ps = torch.zeros(max_len, dtype=s.dtype, device=s.device)
|
| 351 |
+
pm = torch.zeros(max_len, dtype=m.dtype, device=m.device)
|
| 352 |
+
pt = torch.zeros(max_len, dtype=t.dtype, device=t.device)
|
| 353 |
+
ps[: s.size(0)] = s
|
| 354 |
+
pm[: m.size(0)] = m
|
| 355 |
+
pt[: t.size(0)] = t
|
| 356 |
+
padded_seqs.append(ps)
|
| 357 |
+
padded_masks.append(pm)
|
| 358 |
+
padded_tts.append(pt)
|
| 359 |
+
|
| 360 |
+
rid = torch.stack(padded_seqs)
|
| 361 |
+
ram = torch.stack(padded_masks)
|
| 362 |
+
rtt = torch.stack(padded_tts) if token_type_ids is not None else None
|
| 363 |
+
|
| 364 |
+
return rid, ram, rtt, pair_nums, group_indices
|
| 365 |
+
|
| 366 |
+
def _restore_original_order(
|
| 367 |
+
self,
|
| 368 |
+
logits: torch.Tensor,
|
| 369 |
+
group_indices: List[List[int]],
|
| 370 |
+
) -> torch.Tensor:
|
| 371 |
+
"""
|
| 372 |
+
Map flattened logits back so each original index gets its passage score.
|
| 373 |
+
"""
|
| 374 |
+
out = torch.zeros(logits.size(0), dtype=logits.dtype, device=logits.device)
|
| 375 |
+
i = 0
|
| 376 |
+
for indices in group_indices:
|
| 377 |
+
for idx in indices:
|
| 378 |
+
out[idx] = logits[i]
|
| 379 |
+
i += 1
|
| 380 |
+
return out.reshape(-1, 1)
|
| 381 |
|
| 382 |
def average_pooling(self, hidden_state, attention_mask):
|
| 383 |
extended_attention_mask = (
|
|
|
|
| 395 |
cls, model_name_or_path, config: Optional[ListConRankerConfig] = None, **kwargs
|
| 396 |
):
|
| 397 |
model = super().from_pretrained(model_name_or_path, config=config, **kwargs)
|
| 398 |
+
model.hf_model = BertModel.from_pretrained(
|
| 399 |
+
model_name_or_path, config=model.config.bert_config
|
| 400 |
+
)
|
| 401 |
|
| 402 |
+
linear_path = os.path.join(model_name_or_path, "linear_in_embedding.pt")
|
| 403 |
+
transformer_path = os.path.join(model_name_or_path, "list_transformer.pt")
|
|
|
|
| 404 |
|
| 405 |
try:
|
| 406 |
model.linear_in_embedding.load_state_dict(torch.load(linear_path))
|
| 407 |
model.list_transformer.load_state_dict(torch.load(transformer_path))
|
| 408 |
+
except FileNotFoundError as e:
|
| 409 |
+
raise e
|
| 410 |
|
| 411 |
return model
|
| 412 |
+
|
| 413 |
+
def multi_passage(
|
| 414 |
+
self,
|
| 415 |
+
sentences: List[List[str]],
|
| 416 |
+
batch_size: int = 32,
|
| 417 |
+
tokenizer: AutoTokenizer = AutoTokenizer.from_pretrained(
|
| 418 |
+
"ByteDance/ListConRanker"
|
| 419 |
+
),
|
| 420 |
+
):
|
| 421 |
+
"""
|
| 422 |
+
Process multiple passages for each query.
|
| 423 |
+
:param sentences: List of lists, where each inner list contains sentences for a query.
|
| 424 |
+
:return: Tensor of logits for each passage.
|
| 425 |
+
"""
|
| 426 |
+
pairs = []
|
| 427 |
+
for batch in sentences:
|
| 428 |
+
if len(batch) < 2:
|
| 429 |
+
raise ValueError("Each query must have at least one passage.")
|
| 430 |
+
query = batch[0]
|
| 431 |
+
passages = batch[1:]
|
| 432 |
+
for passage in passages:
|
| 433 |
+
pairs.append((query, passage))
|
| 434 |
+
|
| 435 |
+
total_batches = (len(pairs) + batch_size - 1) // batch_size
|
| 436 |
+
total_logits = torch.zeros(len(pairs), dtype=torch.float, device=self.device)
|
| 437 |
+
for batch in range(total_batches):
|
| 438 |
+
batch_pairs = pairs[batch * batch_size : (batch + 1) * batch_size]
|
| 439 |
+
inputs = tokenizer(
|
| 440 |
+
batch_pairs,
|
| 441 |
+
padding=True,
|
| 442 |
+
truncation=True,
|
| 443 |
+
return_tensors="pt",
|
| 444 |
+
)
|
| 445 |
+
logits = self(**inputs)[0]
|
| 446 |
+
total_logits[batch * batch_size : (batch + 1) * batch_size] = (
|
| 447 |
+
logits.squeeze(1)
|
| 448 |
+
)
|
| 449 |
+
return total_logits
|
tokenizer_config.json
CHANGED
|
@@ -47,7 +47,7 @@
|
|
| 47 |
"do_lower_case": true,
|
| 48 |
"mask_token": "[MASK]",
|
| 49 |
"max_length": 512,
|
| 50 |
-
"model_max_length":
|
| 51 |
"never_split": null,
|
| 52 |
"pad_to_multiple_of": null,
|
| 53 |
"pad_token": "[PAD]",
|
|
|
|
| 47 |
"do_lower_case": true,
|
| 48 |
"mask_token": "[MASK]",
|
| 49 |
"max_length": 512,
|
| 50 |
+
"model_max_length": 512,
|
| 51 |
"never_split": null,
|
| 52 |
"pad_to_multiple_of": null,
|
| 53 |
"pad_token": "[PAD]",
|