xiaowenbin commited on
Commit
128ae92
1 Parent(s): d4009b0

init commit

Browse files
.gitattributes CHANGED
@@ -26,3 +26,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
26
  *.zip filter=lfs diff=lfs merge=lfs -text
27
  *.zstandard filter=lfs diff=lfs merge=lfs -text
28
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
26
  *.zip filter=lfs diff=lfs merge=lfs -text
27
  *.zstandard filter=lfs diff=lfs merge=lfs -text
28
  *tfevents* filter=lfs diff=lfs merge=lfs -text
29
+ pytorch_model.bin filter=lfs diff=lfs merge=lfs -text
1_Pooling/config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "word_embedding_dimension": 768,
3
+ "pooling_mode_cls_token": false,
4
+ "pooling_mode_mean_tokens": true,
5
+ "pooling_mode_max_tokens": false,
6
+ "pooling_mode_mean_sqrt_len_tokens": false
7
+ }
README.md ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pipeline_tag: sentence-similarity
3
+ tags:
4
+ - sentence-transformers
5
+ - feature-extraction
6
+ - sentence-similarity
7
+ - transformers
8
+ - semantic-search
9
+ - chinese
10
+ ---
11
+
12
+ # DMetaSoul/sbert-chinese-qmc-domain-v1
13
+
14
+ 此模型是基于之前开源[问题匹配模型](https://huggingface.co/DMetaSoul/sbert-chinese-qmc-domain-v1)的蒸馏轻量化版本(仅含4层 BERT),适用于**开放领域的问题匹配**场景,比如:
15
+
16
+
17
+ - 洗澡用什么香皂好?vs. 洗澡用什么香皂好
18
+ - 大连哪里拍婚纱照好点? vs. 大连哪里拍婚纱照比较好
19
+ - 银行卡怎样挂失?vs. 银行卡丢了怎么挂失啊?
20
+
21
+ 离线训练好的大模型如果直接用于线上推理,对计算资源有苛刻的需求,而且难以满足业务环境对延迟、吞吐量等性能指标的要求,这里我们使用蒸馏手段来把大模型轻量化。从 12 层 BERT 蒸馏为 4 层后,模型参数量缩小到 44%,大概 latency 减半、throughput 翻倍、精度下降 4% 左右(具体结果详见下文评估小节)。
22
+
23
+ # Usage
24
+
25
+ ## 1. Sentence-Transformers
26
+
27
+ 通过 [sentence-transformers](https://www.SBERT.net) 框架来使用该模型,首先进行安装:
28
+
29
+ ```
30
+ pip install -U sentence-transformers
31
+ ```
32
+
33
+ 然后使用下面的代码来载入该模型并进行文本表征向量的提取:
34
+
35
+ ```python
36
+ from sentence_transformers import SentenceTransformer
37
+ sentences = ["我的儿子!他猛然间喊道,我的儿子在哪儿?", "我的儿子呢!他突然喊道,我的儿子在哪里?"]
38
+
39
+ model = SentenceTransformer('DMetaSoul/sbert-chinese-qmc-domain-v1')
40
+ embeddings = model.encode(sentences)
41
+ print(embeddings)
42
+ ```
43
+
44
+ ## 2. HuggingFace Transformers
45
+
46
+ 如果不想使用 [sentence-transformers](https://www.SBERT.net) 的话,也可以通过 HuggingFace Transformers 来载入该模型并进行文本向量抽取:
47
+
48
+ ```python
49
+ from transformers import AutoTokenizer, AutoModel
50
+ import torch
51
+
52
+
53
+ #Mean Pooling - Take attention mask into account for correct averaging
54
+ def mean_pooling(model_output, attention_mask):
55
+ token_embeddings = model_output[0] #First element of model_output contains all token embeddings
56
+ input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
57
+ return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
58
+
59
+
60
+ # Sentences we want sentence embeddings for
61
+ sentences = ["我的儿子!他猛然间喊道,我的儿子在哪儿?", "我的儿子呢!他突然喊道,我的儿子在哪里?"]
62
+
63
+ # Load model from HuggingFace Hub
64
+ tokenizer = AutoTokenizer.from_pretrained('DMetaSoul/sbert-chinese-qmc-domain-v1')
65
+ model = AutoModel.from_pretrained('DMetaSoul/sbert-chinese-qmc-domain-v1')
66
+
67
+ # Tokenize sentences
68
+ encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
69
+
70
+ # Compute token embeddings
71
+ with torch.no_grad():
72
+ model_output = model(**encoded_input)
73
+
74
+ # Perform pooling. In this case, mean pooling.
75
+ sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
76
+
77
+ print("Sentence embeddings:")
78
+ print(sentence_embeddings)
79
+ ```
80
+
81
+ ## Evaluation
82
+
83
+ 这里主要跟蒸馏前对应的 teacher 模型作了对比
84
+
85
+ *性能:*
86
+
87
+ | | Teacher | Student | Gap |
88
+ | ---------- | --------------------- | ------------------- | ----- |
89
+ | Model | BERT-12-layers (102M) | BERT-4-layers (45M) | 0.44x |
90
+ | Cost | 23s | 12s | -47% |
91
+ | Latency | 38ms | 20ms | -47% |
92
+ | Throughput | 421 sentence/s | 791 sentence/s | 1.9x |
93
+
94
+ *精度:*
95
+
96
+ | | **csts_dev** | **csts_test** | **afqmc** | **lcqmc** | **bqcorpus** | **pawsx** | **xiaobu** | **Avg** |
97
+ | -------------- | ------------ | ------------- | --------- | --------- | ------------ | --------- | ---------- | ------- |
98
+ | **Teacher** | 80.90% | 76.62% | 34.51% | 77.05% | 52.95% | 12.97% | 59.47% | 56.35% |
99
+ | **Student** | 79.89% | 76.34% | 27.59% | 69.26% | 49.40% | 9.06% | 53.52% | 52.15% |
100
+ | **Gap** (abs.) | - | - | - | - | - | - | - | -4.2% |
101
+
102
+ *基于1万条数据测试,GPU设备是V100,batch_size=16,max_seq_len=256*
103
+
104
+ ## Citing & Authors
105
+
106
+ E-mail: xiaowenbin@dmetasoul.com
config.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "releases/sbert-chinese-qmc-domain-v1-distill/",
3
+ "architectures": [
4
+ "BertModel"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "classifier_dropout": null,
8
+ "directionality": "bidi",
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 768,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3072,
14
+ "layer_norm_eps": 1e-12,
15
+ "max_position_embeddings": 512,
16
+ "model_type": "bert",
17
+ "num_attention_heads": 12,
18
+ "num_hidden_layers": 4,
19
+ "pad_token_id": 0,
20
+ "pooler_fc_size": 768,
21
+ "pooler_num_attention_heads": 12,
22
+ "pooler_num_fc_layers": 3,
23
+ "pooler_size_per_head": 128,
24
+ "pooler_type": "first_token_transform",
25
+ "position_embedding_type": "absolute",
26
+ "torch_dtype": "float32",
27
+ "transformers_version": "4.16.0",
28
+ "type_vocab_size": 2,
29
+ "use_cache": true,
30
+ "vocab_size": 21128
31
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "sentence_transformers": "2.1.0",
4
+ "transformers": "4.16.0",
5
+ "pytorch": "1.10.2"
6
+ }
7
+ }
modules.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "idx": 0,
4
+ "name": "0",
5
+ "path": "",
6
+ "type": "sentence_transformers.models.Transformer"
7
+ },
8
+ {
9
+ "idx": 1,
10
+ "name": "1",
11
+ "path": "1_Pooling",
12
+ "type": "sentence_transformers.models.Pooling"
13
+ }
14
+ ]
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad505e52d15c2e6c396da1e3ff39c4707368d1c9e91a5a8ed18c18a46e590b24
3
+ size 182288973
sentence_bert_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "max_seq_length": 256,
3
+ "do_lower_case": false
4
+ }
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"do_lower_case": false, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "releases/sbert-chinese-qmc-domain-v1-distill/", "tokenizer_class": "BertTokenizer"}
vocab.txt ADDED
The diff for this file is too large to render. See raw diff