florian-hoenicke commited on
Commit
6e277e4
1 Parent(s): 49f7aec

feat: push custom model

Browse files
1_Pooling/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "word_embedding_dimension": 768,
3
+ "pooling_mode_cls_token": true,
4
+ "pooling_mode_mean_tokens": false,
5
+ "pooling_mode_max_tokens": false,
6
+ "pooling_mode_mean_sqrt_len_tokens": false,
7
+ "pooling_mode_weightedmean_tokens": false,
8
+ "pooling_mode_lasttoken": false,
9
+ "include_prompt": true
10
+ }
README.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ datasets:
4
+ - fine-tuned/jina-embeddings-v2-base-code-5222024-i8af-webapp
5
+ - allenai/c4
6
+ language:
7
+ - en
8
+ pipeline_tag: feature-extraction
9
+ tags:
10
+ - sentence-transformers
11
+ - feature-extraction
12
+ - sentence-similarity
13
+ - mteb
14
+ - Issues
15
+ - Bugs
16
+ - Features
17
+ - Comments
18
+ - Logs
19
+ ---
20
+ This model is a fine-tuned version of [**jinaai/jina-embeddings-v2-base-code**](https://huggingface.co/jinaai/jina-embeddings-v2-base-code) designed for the following use case:
21
+
22
+ issue tracking search
23
+
24
+ ## How to Use
25
+ This model can be easily integrated into your NLP pipeline for tasks such as text classification, sentiment analysis, entity recognition, and more. Here's a simple example to get you started:
26
+
27
+ ```python
28
+ from sentence_transformers import SentenceTransformer
29
+ from sentence_transformers.util import cos_sim
30
+
31
+ model = SentenceTransformer(
32
+ 'fine-tuned/jina-embeddings-v2-base-code-5222024-i8af-webapp',
33
+ trust_remote_code=True
34
+ )
35
+
36
+ embeddings = model.encode([
37
+ 'first text to embed',
38
+ 'second text to embed'
39
+ ])
40
+ print(cos_sim(embeddings[0], embeddings[1]))
41
+ ```
config.json ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "mytmp/finetuned_model",
3
+ "architectures": [
4
+ "JinaBertModel"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.0,
7
+ "attn_implementation": null,
8
+ "auto_map": {
9
+ "AutoConfig": "configuration_bert.JinaBertConfig",
10
+ "AutoModel": "modeling_bert.JinaBertModel",
11
+ "AutoModelForMaskedLM": "jinaai/jina-bert-v2-qk-post-norm--modeling_bert.JinaBertForMaskedLM",
12
+ "AutoModelForSequenceClassification": "jinaai/jina-bert-v2-qk-post-norm--modeling_bert.JinaBertForSequenceClassification"
13
+ },
14
+ "classifier_dropout": null,
15
+ "emb_pooler": "mean",
16
+ "feed_forward_type": "geglu",
17
+ "gradient_checkpointing": false,
18
+ "hidden_act": "gelu",
19
+ "hidden_dropout_prob": 0.0,
20
+ "hidden_size": 768,
21
+ "initializer_range": 0.02,
22
+ "intermediate_size": 3072,
23
+ "layer_norm_eps": 1e-12,
24
+ "max_position_embeddings": 8192,
25
+ "model_max_length": 8192,
26
+ "model_type": "bert",
27
+ "num_attention_heads": 12,
28
+ "num_hidden_layers": 12,
29
+ "pad_token_id": 0,
30
+ "position_embedding_type": "alibi",
31
+ "torch_dtype": "float32",
32
+ "transformers_version": "4.40.2",
33
+ "type_vocab_size": 2,
34
+ "use_cache": true,
35
+ "vocab_size": 61056
36
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "sentence_transformers": "2.7.0",
4
+ "transformers": "4.40.2",
5
+ "pytorch": "2.3.0+cu121"
6
+ },
7
+ "prompts": {},
8
+ "default_prompt_name": null
9
+ }
configuration_bert.py ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ # Copyright (c) 2023 Jina AI GmbH. All rights reserved.
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+ """ BERT model configuration"""
18
+ from collections import OrderedDict
19
+ from typing import Mapping
20
+ import warnings
21
+
22
+ from transformers.configuration_utils import PretrainedConfig
23
+ from transformers.utils import logging
24
+
25
+ try:
26
+ from optimum.exporters.onnx.model_configs import BertOnnxConfig
27
+ OPTIMUM_INSTALLED = True
28
+ except ImportError:
29
+ warnings.warn("optimum is not installed. To use OnnxConfig and BertOnnxConfig, make sure that `optimum` package is installed")
30
+ OPTIMUM_INSTALLED = False
31
+
32
+
33
+ logger = logging.get_logger(__name__)
34
+
35
+
36
+ class JinaBertConfig(PretrainedConfig):
37
+ r"""
38
+ This is the configuration class to store the configuration of a [`JinaBertModel`]. It is used to
39
+ instantiate a BERT model according to the specified arguments, defining the model architecture. Instantiating a
40
+ configuration with the defaults will yield a similar configuration to that of the BERT
41
+ [bert-base-uncased](https://huggingface.co/bert-base-uncased) architecture.
42
+
43
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
44
+ documentation from [`PretrainedConfig`] for more information.
45
+
46
+
47
+ Args:
48
+ vocab_size (`int`, *optional*, defaults to 30522):
49
+ Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
50
+ `inputs_ids` passed when calling [`BertModel`] or [`TFBertModel`].
51
+ hidden_size (`int`, *optional*, defaults to 768):
52
+ Dimensionality of the encoder layers and the pooler layer.
53
+ num_hidden_layers (`int`, *optional*, defaults to 12):
54
+ Number of hidden layers in the Transformer encoder.
55
+ num_attention_heads (`int`, *optional*, defaults to 12):
56
+ Number of attention heads for each attention layer in the Transformer encoder.
57
+ intermediate_size (`int`, *optional*, defaults to 3072):
58
+ Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
59
+ hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
60
+ The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
61
+ `"relu"`, `"silu"` and `"gelu_new"` are supported.
62
+ hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
63
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
64
+ attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
65
+ The dropout ratio for the attention probabilities.
66
+ max_position_embeddings (`int`, *optional*, defaults to 512):
67
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
68
+ just in case (e.g., 512 or 1024 or 2048).
69
+ type_vocab_size (`int`, *optional*, defaults to 2):
70
+ The vocabulary size of the `token_type_ids` passed when calling [`BertModel`] or [`TFBertModel`].
71
+ initializer_range (`float`, *optional*, defaults to 0.02):
72
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
73
+ layer_norm_eps (`float`, *optional*, defaults to 1e-12):
74
+ The epsilon used by the layer normalization layers.
75
+ position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
76
+ Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
77
+ positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
78
+ [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
79
+ For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
80
+ with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
81
+ is_decoder (`bool`, *optional*, defaults to `False`):
82
+ Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
83
+ use_cache (`bool`, *optional*, defaults to `True`):
84
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
85
+ relevant if `config.is_decoder=True`.
86
+ classifier_dropout (`float`, *optional*):
87
+ The dropout ratio for the classification head.
88
+ feed_forward_type (`str`, *optional*, defaults to `"original"`):
89
+ The type of feed forward layer to use in the bert layers.
90
+ Can be one of GLU variants, e.g. `"reglu"`, `"geglu"`
91
+ emb_pooler (`str`, *optional*, defaults to `None`):
92
+ The function to use for pooling the last layer embeddings to get the sentence embeddings.
93
+ Should be one of `None`, `"mean"`.
94
+ attn_implementation (`str`, *optional*, defaults to `"torch"`):
95
+ The implementation of the self-attention layer. Can be one of:
96
+ - `None` for the original implementation,
97
+ - `torch` for the PyTorch SDPA implementation,
98
+
99
+ Examples:
100
+
101
+ ```python
102
+ >>> from transformers import JinaBertConfig, JinaBertModel
103
+
104
+ >>> # Initializing a JinaBert configuration
105
+ >>> configuration = JinaBertConfig()
106
+
107
+ >>> # Initializing a model (with random weights) from the configuration
108
+ >>> model = JinaBertModel(configuration)
109
+
110
+ >>> # Accessing the model configuration
111
+ >>> configuration = model.config
112
+
113
+ >>> # Encode text inputs
114
+ >>> embeddings = model.encode(text_inputs)
115
+ ```"""
116
+ model_type = "bert"
117
+
118
+ def __init__(
119
+ self,
120
+ vocab_size=30522,
121
+ hidden_size=768,
122
+ num_hidden_layers=12,
123
+ num_attention_heads=12,
124
+ intermediate_size=3072,
125
+ hidden_act="gelu",
126
+ hidden_dropout_prob=0.1,
127
+ attention_probs_dropout_prob=0.1,
128
+ max_position_embeddings=512,
129
+ type_vocab_size=2,
130
+ initializer_range=0.02,
131
+ layer_norm_eps=1e-12,
132
+ pad_token_id=0,
133
+ position_embedding_type="absolute",
134
+ use_cache=True,
135
+ classifier_dropout=None,
136
+ feed_forward_type="original",
137
+ emb_pooler=None,
138
+ attn_implementation=None,
139
+ **kwargs,
140
+ ):
141
+ super().__init__(pad_token_id=pad_token_id, **kwargs)
142
+
143
+ self.vocab_size = vocab_size
144
+ self.hidden_size = hidden_size
145
+ self.num_hidden_layers = num_hidden_layers
146
+ self.num_attention_heads = num_attention_heads
147
+ self.hidden_act = hidden_act
148
+ self.intermediate_size = intermediate_size
149
+ self.hidden_dropout_prob = hidden_dropout_prob
150
+ self.attention_probs_dropout_prob = attention_probs_dropout_prob
151
+ self.max_position_embeddings = max_position_embeddings
152
+ self.type_vocab_size = type_vocab_size
153
+ self.initializer_range = initializer_range
154
+ self.layer_norm_eps = layer_norm_eps
155
+ self.position_embedding_type = position_embedding_type
156
+ self.use_cache = use_cache
157
+ self.classifier_dropout = classifier_dropout
158
+ self.feed_forward_type = feed_forward_type
159
+ self.emb_pooler = emb_pooler
160
+ self.attn_implementation = attn_implementation
161
+
162
+ if OPTIMUM_INSTALLED:
163
+
164
+ class JinaBertOnnxConfig(BertOnnxConfig):
165
+
166
+ @property
167
+ def inputs(self) -> Mapping[str, Mapping[int, str]]:
168
+ if self.task == "multiple-choice":
169
+ dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
170
+ else:
171
+ dynamic_axis = {0: "batch", 1: "sequence"}
172
+ return OrderedDict(
173
+ [
174
+ ("input_ids", dynamic_axis),
175
+ ("attention_mask", dynamic_axis),
176
+ ]
177
+ )
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ebf1dc39f8cbbb0d9725b7e7e77619ec880339d422a8f1bf71c6c3700f658dfa
3
+ size 643505600
modeling_bert.py ADDED
@@ -0,0 +1,2312 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ # Copyright (c) 2023 Jina AI GmbH. All rights reserved.
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+ """PyTorch BERT model."""
18
+
19
+
20
+ import math
21
+ import os
22
+ import warnings
23
+ from dataclasses import dataclass
24
+ from typing import List, Optional, Tuple, Union
25
+ import numpy as np
26
+
27
+ import torch
28
+ import torch.utils.checkpoint
29
+ from torch import nn
30
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
31
+
32
+ from transformers.activations import ACT2FN
33
+ from transformers.modeling_outputs import (
34
+ BaseModelOutputWithPastAndCrossAttentions,
35
+ BaseModelOutputWithPoolingAndCrossAttentions,
36
+ CausalLMOutputWithCrossAttentions,
37
+ MaskedLMOutput,
38
+ MultipleChoiceModelOutput,
39
+ NextSentencePredictorOutput,
40
+ QuestionAnsweringModelOutput,
41
+ SequenceClassifierOutput,
42
+ TokenClassifierOutput,
43
+ )
44
+ from transformers.modeling_utils import PreTrainedModel
45
+ from transformers.pytorch_utils import (
46
+ apply_chunking_to_forward,
47
+ find_pruneable_heads_and_indices,
48
+ prune_linear_layer,
49
+ )
50
+ from transformers.utils import (
51
+ ModelOutput,
52
+ add_code_sample_docstrings,
53
+ add_start_docstrings,
54
+ add_start_docstrings_to_model_forward,
55
+ logging,
56
+ replace_return_docstrings,
57
+ )
58
+ from .configuration_bert import JinaBertConfig
59
+
60
+ # Torch implementation
61
+ try:
62
+ from torch.nn.functional import scaled_dot_product_attention
63
+ except ImportError:
64
+ scaled_dot_product_attention = None
65
+
66
+ # This is used by encode but user may not have it installed
67
+ try:
68
+ from tqdm.autonotebook import trange
69
+
70
+ has_tqdm = True
71
+ except ImportError:
72
+ has_tqdm = False
73
+
74
+ logger = logging.get_logger(__name__)
75
+
76
+ _CHECKPOINT_FOR_DOC = "bert-base-uncased"
77
+ _CONFIG_FOR_DOC = "JinaBertConfig"
78
+
79
+ # TokenClassification docstring
80
+ _CHECKPOINT_FOR_TOKEN_CLASSIFICATION = (
81
+ "dbmdz/bert-large-cased-finetuned-conll03-english"
82
+ )
83
+ _TOKEN_CLASS_EXPECTED_OUTPUT = "['O', 'I-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O', 'O', 'I-LOC', 'O', 'I-LOC', 'I-LOC'] "
84
+ _TOKEN_CLASS_EXPECTED_LOSS = 0.01
85
+
86
+ # QuestionAnswering docstring
87
+ _CHECKPOINT_FOR_QA = "deepset/bert-base-cased-squad2"
88
+ _QA_EXPECTED_OUTPUT = "'a nice puppet'"
89
+ _QA_EXPECTED_LOSS = 7.41
90
+ _QA_TARGET_START_INDEX = 14
91
+ _QA_TARGET_END_INDEX = 15
92
+
93
+ # SequenceClassification docstring
94
+ _CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION = "textattack/bert-base-uncased-yelp-polarity"
95
+ _SEQ_CLASS_EXPECTED_OUTPUT = "'LABEL_1'"
96
+ _SEQ_CLASS_EXPECTED_LOSS = 0.01
97
+
98
+
99
+ def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
100
+ """Load tf checkpoints in a pytorch model."""
101
+ try:
102
+ import re
103
+
104
+ import numpy as np
105
+ import tensorflow as tf
106
+ except ImportError:
107
+ logger.error(
108
+ "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
109
+ "https://www.tensorflow.org/install/ for installation instructions."
110
+ )
111
+ raise
112
+ tf_path = os.path.abspath(tf_checkpoint_path)
113
+ logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
114
+ # Load weights from TF model
115
+ init_vars = tf.train.list_variables(tf_path)
116
+ names = []
117
+ arrays = []
118
+ for name, shape in init_vars:
119
+ logger.info(f"Loading TF weight {name} with shape {shape}")
120
+ array = tf.train.load_variable(tf_path, name)
121
+ names.append(name)
122
+ arrays.append(array)
123
+
124
+ for name, array in zip(names, arrays):
125
+ name = name.split("/")
126
+ # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
127
+ # which are not required for using pretrained model
128
+ if any(
129
+ n
130
+ in [
131
+ "adam_v",
132
+ "adam_m",
133
+ "AdamWeightDecayOptimizer",
134
+ "AdamWeightDecayOptimizer_1",
135
+ "global_step",
136
+ ]
137
+ for n in name
138
+ ):
139
+ logger.info(f"Skipping {'/'.join(name)}")
140
+ continue
141
+ pointer = model
142
+ for m_name in name:
143
+ if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
144
+ scope_names = re.split(r"_(\d+)", m_name)
145
+ else:
146
+ scope_names = [m_name]
147
+ if scope_names[0] == "kernel" or scope_names[0] == "gamma":
148
+ pointer = getattr(pointer, "weight")
149
+ elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
150
+ pointer = getattr(pointer, "bias")
151
+ elif scope_names[0] == "output_weights":
152
+ pointer = getattr(pointer, "weight")
153
+ elif scope_names[0] == "squad":
154
+ pointer = getattr(pointer, "classifier")
155
+ else:
156
+ try:
157
+ pointer = getattr(pointer, scope_names[0])
158
+ except AttributeError:
159
+ logger.info(f"Skipping {'/'.join(name)}")
160
+ continue
161
+ if len(scope_names) >= 2:
162
+ num = int(scope_names[1])
163
+ pointer = pointer[num]
164
+ if m_name[-11:] == "_embeddings":
165
+ pointer = getattr(pointer, "weight")
166
+ elif m_name == "kernel":
167
+ array = np.transpose(array)
168
+ try:
169
+ if pointer.shape != array.shape:
170
+ raise ValueError(
171
+ f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
172
+ )
173
+ except ValueError as e:
174
+ e.args += (pointer.shape, array.shape)
175
+ raise
176
+ logger.info(f"Initialize PyTorch weight {name}")
177
+ pointer.data = torch.from_numpy(array)
178
+ return model
179
+
180
+
181
+ class JinaBertEmbeddings(nn.Module):
182
+ """Construct the embeddings from word, position and token_type embeddings."""
183
+
184
+ def __init__(self, config: JinaBertConfig):
185
+ super().__init__()
186
+ self.word_embeddings = nn.Embedding(
187
+ config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
188
+ )
189
+ if config.position_embedding_type != "alibi":
190
+ self.position_embeddings = nn.Embedding(
191
+ config.max_position_embeddings, config.hidden_size
192
+ )
193
+ self.token_type_embeddings = nn.Embedding(
194
+ config.type_vocab_size, config.hidden_size
195
+ )
196
+
197
+ # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
198
+ # any TensorFlow checkpoint file
199
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
200
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
201
+ # position_ids (1, len position emb) is contiguous in memory and exported when serialized
202
+ self.position_embedding_type = getattr(
203
+ config, "position_embedding_type", "absolute"
204
+ )
205
+ self.register_buffer(
206
+ "position_ids",
207
+ torch.arange(config.max_position_embeddings).expand((1, -1)),
208
+ persistent=False,
209
+ )
210
+ self.register_buffer(
211
+ "token_type_ids",
212
+ torch.zeros(self.position_ids.size(), dtype=torch.long),
213
+ persistent=False,
214
+ )
215
+
216
+ def forward(
217
+ self,
218
+ input_ids: Optional[torch.LongTensor] = None,
219
+ token_type_ids: Optional[torch.LongTensor] = None,
220
+ position_ids: Optional[torch.LongTensor] = None,
221
+ inputs_embeds: Optional[torch.FloatTensor] = None,
222
+ past_key_values_length: int = 0,
223
+ ) -> torch.Tensor:
224
+ if input_ids is not None:
225
+ input_shape = input_ids.size()
226
+ else:
227
+ input_shape = inputs_embeds.size()[:-1]
228
+
229
+ seq_length = input_shape[1]
230
+
231
+ if position_ids is None:
232
+ position_ids = self.position_ids[
233
+ :, past_key_values_length : seq_length + past_key_values_length
234
+ ]
235
+
236
+ # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
237
+ # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
238
+ # issue #5664
239
+ if token_type_ids is None:
240
+ if hasattr(self, "token_type_ids"):
241
+ buffered_token_type_ids = self.token_type_ids[:, :seq_length]
242
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
243
+ input_shape[0], seq_length
244
+ )
245
+ token_type_ids = buffered_token_type_ids_expanded
246
+ else:
247
+ token_type_ids = torch.zeros(
248
+ input_shape, dtype=torch.long, device=self.position_ids.device
249
+ )
250
+
251
+ if inputs_embeds is None:
252
+ inputs_embeds = self.word_embeddings(input_ids)
253
+ token_type_embeddings = self.token_type_embeddings(token_type_ids)
254
+
255
+ embeddings = inputs_embeds + token_type_embeddings
256
+ if self.position_embedding_type == "absolute":
257
+ position_embeddings = self.position_embeddings(position_ids)
258
+ embeddings += position_embeddings
259
+ embeddings = self.LayerNorm(embeddings)
260
+ embeddings = self.dropout(embeddings)
261
+ return embeddings
262
+
263
+
264
+ class JinaBertSelfAttention(nn.Module):
265
+ def __init__(self, config: JinaBertConfig, position_embedding_type=None):
266
+ super().__init__()
267
+ if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
268
+ config, "embedding_size"
269
+ ):
270
+ raise ValueError(
271
+ f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
272
+ f"heads ({config.num_attention_heads})"
273
+ )
274
+
275
+ self.attn_implementation = config.attn_implementation
276
+ self.num_attention_heads = config.num_attention_heads
277
+ self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
278
+ self.all_head_size = self.num_attention_heads * self.attention_head_size
279
+
280
+ self.query = nn.Linear(config.hidden_size, self.all_head_size)
281
+ self.key = nn.Linear(config.hidden_size, self.all_head_size)
282
+ self.value = nn.Linear(config.hidden_size, self.all_head_size)
283
+ self.layer_norm_q = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
284
+ self.layer_norm_k = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
285
+
286
+ self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
287
+ self.position_embedding_type = position_embedding_type or getattr(
288
+ config, "position_embedding_type", "absolute"
289
+ )
290
+ if (
291
+ self.position_embedding_type == "relative_key"
292
+ or self.position_embedding_type == "relative_key_query"
293
+ ):
294
+ self.max_position_embeddings = config.max_position_embeddings
295
+ self.distance_embedding = nn.Embedding(
296
+ 2 * config.max_position_embeddings - 1, self.attention_head_size
297
+ )
298
+
299
+ self.is_decoder = config.is_decoder
300
+
301
+ def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
302
+ new_x_shape = x.size()[:-1] + (
303
+ self.num_attention_heads,
304
+ self.attention_head_size,
305
+ )
306
+ x = x.view(new_x_shape)
307
+ return x.permute(0, 2, 1, 3)
308
+
309
+ def forward(
310
+ self,
311
+ hidden_states: torch.Tensor,
312
+ attention_mask: Optional[torch.FloatTensor] = None,
313
+ head_mask: Optional[torch.FloatTensor] = None,
314
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
315
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
316
+ past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
317
+ output_attentions: Optional[bool] = False,
318
+ bias: Optional[torch.FloatTensor] = None,
319
+ ) -> Tuple[torch.Tensor]:
320
+ mixed_query_layer = self.layer_norm_q(self.query(hidden_states))
321
+
322
+ # If this is instantiated as a cross-attention module, the keys
323
+ # and values come from an encoder; the attention mask needs to be
324
+ # such that the encoder's padding tokens are not attended to.
325
+ is_cross_attention = encoder_hidden_states is not None
326
+
327
+ if is_cross_attention and past_key_value is not None:
328
+ # reuse k,v, cross_attentions
329
+ key_layer = past_key_value[0]
330
+ value_layer = past_key_value[1]
331
+ attention_mask = encoder_attention_mask
332
+ elif is_cross_attention:
333
+ key_layer = self.transpose_for_scores(self.layer_norm_k(self.key(encoder_hidden_states)))
334
+ value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
335
+ attention_mask = encoder_attention_mask
336
+ elif past_key_value is not None:
337
+ key_layer = self.transpose_for_scores(self.layer_norm_k(self.key(hidden_states)))
338
+ value_layer = self.transpose_for_scores(self.value(hidden_states))
339
+ key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
340
+ value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
341
+ else:
342
+ key_layer = self.transpose_for_scores(self.layer_norm_k(self.key(hidden_states)))
343
+ value_layer = self.transpose_for_scores(self.value(hidden_states))
344
+
345
+ query_layer = self.transpose_for_scores(mixed_query_layer)
346
+
347
+ use_cache = past_key_value is not None
348
+ if self.is_decoder:
349
+ # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
350
+ # Further calls to cross_attention layer can then reuse all cross-attention
351
+ # key/value_states (first "if" case)
352
+ # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
353
+ # all previous decoder key/value_states. Further calls to uni-directional self-attention
354
+ # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
355
+ # if encoder bi-directional self-attention `past_key_value` is always `None`
356
+ past_key_value = (key_layer, value_layer)
357
+
358
+ if self.attn_implementation == 'torch' and scaled_dot_product_attention is not None:
359
+ b, _, s, _ = query_layer.shape
360
+ new_bias = attention_mask + bias
361
+ attn = scaled_dot_product_attention(query_layer, key_layer, value_layer, new_bias)
362
+ attn = attn.permute(0, 2, 1, 3).contiguous()
363
+ return (attn.view(b, s, self.all_head_size),)
364
+
365
+ # Take the dot product between "query" and "key" to get the raw attention scores.
366
+ attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
367
+
368
+ if (
369
+ self.position_embedding_type == "relative_key"
370
+ or self.position_embedding_type == "relative_key_query"
371
+ ):
372
+ query_length, key_length = query_layer.shape[2], key_layer.shape[2]
373
+ if use_cache:
374
+ position_ids_l = torch.tensor(
375
+ key_length - 1, dtype=torch.long, device=hidden_states.device
376
+ ).view(-1, 1)
377
+ else:
378
+ position_ids_l = torch.arange(
379
+ query_length, dtype=torch.long, device=hidden_states.device
380
+ ).view(-1, 1)
381
+ position_ids_r = torch.arange(
382
+ key_length, dtype=torch.long, device=hidden_states.device
383
+ ).view(1, -1)
384
+ distance = position_ids_l - position_ids_r
385
+
386
+ positional_embedding = self.distance_embedding(
387
+ distance + self.max_position_embeddings - 1
388
+ )
389
+ positional_embedding = positional_embedding.to(
390
+ dtype=query_layer.dtype
391
+ ) # fp16 compatibility
392
+
393
+ if self.position_embedding_type == "relative_key":
394
+ relative_position_scores = torch.einsum(
395
+ "bhld,lrd->bhlr", query_layer, positional_embedding
396
+ )
397
+ attention_scores = attention_scores + relative_position_scores
398
+ elif self.position_embedding_type == "relative_key_query":
399
+ relative_position_scores_query = torch.einsum(
400
+ "bhld,lrd->bhlr", query_layer, positional_embedding
401
+ )
402
+ relative_position_scores_key = torch.einsum(
403
+ "bhrd,lrd->bhlr", key_layer, positional_embedding
404
+ )
405
+ attention_scores = (
406
+ attention_scores
407
+ + relative_position_scores_query
408
+ + relative_position_scores_key
409
+ )
410
+
411
+ attention_scores = attention_scores / math.sqrt(self.attention_head_size)
412
+ if attention_mask is not None:
413
+ # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
414
+ attention_scores = attention_scores + attention_mask
415
+
416
+ # Normalize the attention scores to probabilities.
417
+ attention_probs = nn.functional.softmax(attention_scores + bias, dim=-1)
418
+
419
+ # This is actually dropping out entire tokens to attend to, which might
420
+ # seem a bit unusual, but is taken from the original Transformer paper.
421
+ attention_probs = self.dropout(attention_probs)
422
+
423
+ # Mask heads if we want to
424
+ if head_mask is not None:
425
+ attention_probs = attention_probs * head_mask
426
+
427
+ context_layer = torch.matmul(attention_probs, value_layer)
428
+
429
+ context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
430
+ new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
431
+ context_layer = context_layer.view(new_context_layer_shape)
432
+
433
+ outputs = (
434
+ (context_layer, attention_scores) if output_attentions else (context_layer,)
435
+ )
436
+
437
+ if self.is_decoder:
438
+ outputs = outputs + (past_key_value,)
439
+ return outputs
440
+
441
+
442
+ class JinaBertSelfOutput(nn.Module):
443
+ def __init__(self, config):
444
+ super().__init__()
445
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
446
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
447
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
448
+
449
+ def forward(
450
+ self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
451
+ ) -> torch.Tensor:
452
+ hidden_states = self.dense(hidden_states)
453
+ hidden_states = self.dropout(hidden_states)
454
+ hidden_states = self.LayerNorm(hidden_states + input_tensor)
455
+ return hidden_states
456
+
457
+
458
+ class JinaBertAttention(nn.Module):
459
+ def __init__(self, config, position_embedding_type=None):
460
+ super().__init__()
461
+ self.self = JinaBertSelfAttention(
462
+ config, position_embedding_type=position_embedding_type
463
+ )
464
+ self.output = JinaBertSelfOutput(config)
465
+ self.pruned_heads = set()
466
+
467
+ def prune_heads(self, heads):
468
+ if len(heads) == 0:
469
+ return
470
+ heads, index = find_pruneable_heads_and_indices(
471
+ heads,
472
+ self.self.num_attention_heads,
473
+ self.self.attention_head_size,
474
+ self.pruned_heads,
475
+ )
476
+
477
+ # Prune linear layers
478
+ self.self.query = prune_linear_layer(self.self.query, index)
479
+ self.self.key = prune_linear_layer(self.self.key, index)
480
+ self.self.value = prune_linear_layer(self.self.value, index)
481
+ self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
482
+
483
+ # Update hyper params and store pruned heads
484
+ self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
485
+ self.self.all_head_size = (
486
+ self.self.attention_head_size * self.self.num_attention_heads
487
+ )
488
+ self.pruned_heads = self.pruned_heads.union(heads)
489
+
490
+ def forward(
491
+ self,
492
+ hidden_states: torch.Tensor,
493
+ attention_mask: Optional[torch.FloatTensor] = None,
494
+ head_mask: Optional[torch.FloatTensor] = None,
495
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
496
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
497
+ past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
498
+ output_attentions: Optional[bool] = False,
499
+ bias: Optional[torch.FloatTensor] = None,
500
+ ) -> Tuple[torch.Tensor]:
501
+ self_outputs = self.self(
502
+ hidden_states,
503
+ attention_mask,
504
+ head_mask,
505
+ encoder_hidden_states,
506
+ encoder_attention_mask,
507
+ past_key_value,
508
+ output_attentions,
509
+ bias,
510
+ )
511
+ attention_output = self.output(self_outputs[0], hidden_states)
512
+ outputs = (attention_output,) + self_outputs[
513
+ 1:
514
+ ] # add attentions if we output them
515
+ return outputs
516
+
517
+
518
+ class JinaBertMLP(nn.Module):
519
+ def __init__(self, config: JinaBertConfig):
520
+ super().__init__()
521
+ self.config = config
522
+ self.act = ACT2FN[config.hidden_act]
523
+ self.up_layer = nn.Linear(
524
+ config.hidden_size, config.intermediate_size, bias=False
525
+ )
526
+ self.down_layer = nn.Linear(config.intermediate_size, config.hidden_size)
527
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
528
+
529
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
530
+ # Up
531
+ hidden_mlp_states = self.act(self.up_layer(hidden_states))
532
+ hidden_mlp_states = self.dropout(hidden_mlp_states)
533
+ # Down
534
+ return self.down_layer(hidden_mlp_states)
535
+
536
+
537
+ class JinaBertGLUMLP(nn.Module):
538
+ def __init__(self, config: JinaBertConfig):
539
+ super().__init__()
540
+ self.config = config
541
+ if config.feed_forward_type == 'reglu':
542
+ self.act = nn.ReLU()
543
+ elif config.feed_forward_type == 'geglu':
544
+ self.act = nn.GELU()
545
+ else:
546
+ raise ValueError(
547
+ f"feed_forward_type {config.feed_forward_type} not supported"
548
+ )
549
+ self.up_gated_layer = nn.Linear(
550
+ config.hidden_size, config.intermediate_size * 2, bias=False
551
+ )
552
+ self.down_layer = nn.Linear(config.intermediate_size, config.hidden_size)
553
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
554
+
555
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
556
+ # Up with gate
557
+ hidden_mlp_states = self.up_gated_layer(hidden_states)
558
+ up = hidden_mlp_states[:, :, :self.config.intermediate_size]
559
+ gated = hidden_mlp_states[:, :, self.config.intermediate_size:]
560
+ hidden_mlp_states = up * self.act(gated)
561
+ hidden_mlp_states = self.dropout(hidden_mlp_states)
562
+ # Down
563
+ return self.down_layer(hidden_mlp_states)
564
+
565
+
566
+ class JinaBertLayer(nn.Module):
567
+ def __init__(self, config: JinaBertConfig):
568
+ super().__init__()
569
+ self.chunk_size_feed_forward = config.chunk_size_feed_forward
570
+ self.seq_len_dim = 1
571
+ self.attention = JinaBertAttention(config)
572
+ self.is_decoder = config.is_decoder
573
+ self.add_cross_attention = config.add_cross_attention
574
+ self.feed_forward_type = config.feed_forward_type
575
+ self.layer_norm_1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
576
+ self.layer_norm_2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
577
+ if self.add_cross_attention:
578
+ if not self.is_decoder:
579
+ raise ValueError(
580
+ f"{self} should be used as a decoder model if cross attention is added"
581
+ )
582
+ self.crossattention = JinaBertAttention(
583
+ config, position_embedding_type="absolute"
584
+ )
585
+ if self.feed_forward_type.endswith('glu'):
586
+ self.mlp = JinaBertGLUMLP(config)
587
+ else:
588
+ self.mlp = JinaBertMLP(config)
589
+
590
+ def forward(
591
+ self,
592
+ hidden_states: torch.Tensor,
593
+ attention_mask: Optional[torch.FloatTensor] = None,
594
+ head_mask: Optional[torch.FloatTensor] = None,
595
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
596
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
597
+ bias: Optional[torch.FloatTensor] = None,
598
+ past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
599
+ output_attentions: Optional[bool] = False,
600
+ ) -> Tuple[torch.Tensor]:
601
+ # Pre-Norm
602
+ residual = hidden_states
603
+
604
+ # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
605
+ self_attn_past_key_value = (
606
+ past_key_value[:2] if past_key_value is not None else None
607
+ )
608
+ self_attention_outputs = self.attention(
609
+ hidden_states,
610
+ attention_mask,
611
+ head_mask,
612
+ output_attentions=output_attentions,
613
+ past_key_value=self_attn_past_key_value,
614
+ bias=bias,
615
+ )
616
+ attention_output = self_attention_outputs[0]
617
+
618
+ # if decoder, the last output is tuple of self-attn cache
619
+ if self.is_decoder:
620
+ outputs = self_attention_outputs[1:-1]
621
+ present_key_value = self_attention_outputs[-1]
622
+ else:
623
+ outputs = self_attention_outputs[
624
+ 1:
625
+ ] # add self attentions if we output attention weights
626
+
627
+ cross_attn_present_key_value = None
628
+ if self.is_decoder and encoder_hidden_states is not None:
629
+ if not hasattr(self, "crossattention"):
630
+ raise ValueError(
631
+ f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
632
+ " by setting `config.add_cross_attention=True`"
633
+ )
634
+
635
+ # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
636
+ cross_attn_past_key_value = (
637
+ past_key_value[-2:] if past_key_value is not None else None
638
+ )
639
+ cross_attention_outputs = self.crossattention(
640
+ attention_output,
641
+ attention_mask,
642
+ head_mask,
643
+ encoder_hidden_states,
644
+ encoder_attention_mask,
645
+ cross_attn_past_key_value,
646
+ output_attentions,
647
+ )
648
+ attention_output = cross_attention_outputs[0]
649
+ outputs = (
650
+ outputs + cross_attention_outputs[1:-1]
651
+ ) # add cross attentions if we output attention weights
652
+
653
+ # add cross-attn cache to positions 3,4 of present_key_value tuple
654
+ cross_attn_present_key_value = cross_attention_outputs[-1]
655
+ present_key_value = present_key_value + cross_attn_present_key_value
656
+
657
+ residual = self.layer_norm_1(residual + attention_output)
658
+ mlp_output = self.mlp(residual)
659
+ layer_output = self.layer_norm_2(residual + mlp_output)
660
+ outputs = (layer_output,) + outputs
661
+
662
+ # if decoder, return the attn key/values as the last output
663
+ if self.is_decoder:
664
+ outputs = outputs + (present_key_value,)
665
+
666
+ return outputs
667
+
668
+
669
+ class JinaBertEncoder(nn.Module):
670
+ def __init__(self, config: JinaBertConfig):
671
+ super().__init__()
672
+ self.config = config
673
+ self.layer = nn.ModuleList(
674
+ [JinaBertLayer(config) for _ in range(config.num_hidden_layers)]
675
+ )
676
+ self.gradient_checkpointing = False
677
+ self.num_attention_heads = config.num_attention_heads
678
+
679
+ def rebuild_alibi_tensor(
680
+ self, size: int, device: Optional[Union[torch.device, str]] = None
681
+ ):
682
+ # Alibi
683
+ # Following https://github.com/ofirpress/attention_with_linear_biases/issues/5 (Implementation 1)
684
+ # In the causal case, you can exploit the fact that softmax is invariant to a uniform translation
685
+ # of the logits, which makes the math work out *after* applying causal masking. If no causal masking
686
+ # will be applied, it is necessary to construct the diagonal mask.
687
+ n_heads = self.num_attention_heads
688
+
689
+ def _get_alibi_head_slopes(n_heads: int) -> List[float]:
690
+ def get_slopes_power_of_2(n):
691
+ start = 2 ** (-(2 ** -(math.log2(n) - 3)))
692
+ ratio = start
693
+ return [start * ratio**i for i in range(n)]
694
+
695
+ if math.log2(n_heads).is_integer():
696
+ return get_slopes_power_of_2(
697
+ n_heads
698
+ ) # In the paper, we only train models that have 2^a heads for some a. This function has
699
+ else: # some good properties that only occur when the input is a power of 2. To maintain that even
700
+ closest_power_of_2 = 2 ** math.floor(
701
+ math.log2(n_heads)
702
+ ) # when the number of heads is not a power of 2, we use this workaround.
703
+ return (
704
+ get_slopes_power_of_2(closest_power_of_2)
705
+ + _get_alibi_head_slopes(2 * closest_power_of_2)[0::2][
706
+ : n_heads - closest_power_of_2
707
+ ]
708
+ )
709
+
710
+ context_position = torch.arange(size, device=device)[:, None]
711
+ memory_position = torch.arange(size, device=device)[None, :]
712
+ relative_position = torch.abs(memory_position - context_position)
713
+ # [n_heads, max_token_length, max_token_length]
714
+ relative_position = relative_position.unsqueeze(0).expand(n_heads, -1, -1)
715
+ slopes = torch.Tensor(_get_alibi_head_slopes(n_heads)).to(device) * -1
716
+ alibi = slopes.unsqueeze(1).unsqueeze(1) * relative_position
717
+ # [1, n_heads, max_token_length, max_token_length]
718
+ alibi = alibi.unsqueeze(0)
719
+ assert alibi.shape == torch.Size([1, n_heads, size, size])
720
+
721
+ self._current_alibi_size = size
722
+ return alibi
723
+
724
+ def forward(
725
+ self,
726
+ hidden_states: torch.Tensor,
727
+ attention_mask: Optional[torch.FloatTensor] = None,
728
+ head_mask: Optional[torch.FloatTensor] = None,
729
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
730
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
731
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
732
+ use_cache: Optional[bool] = None,
733
+ output_attentions: Optional[bool] = False,
734
+ output_hidden_states: Optional[bool] = False,
735
+ return_dict: Optional[bool] = True,
736
+ ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
737
+ all_hidden_states = () if output_hidden_states else None
738
+ all_self_attentions = () if output_attentions else None
739
+ all_cross_attentions = (
740
+ () if output_attentions and self.config.add_cross_attention else None
741
+ )
742
+
743
+ # Add alibi matrix to extended_attention_mask
744
+ _, seqlen, _ = hidden_states.size()
745
+ alibi_bias = self.rebuild_alibi_tensor(size=seqlen, device=hidden_states.device).to(hidden_states.dtype)
746
+ if self.gradient_checkpointing and self.training:
747
+ if use_cache:
748
+ logger.warning_once(
749
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
750
+ )
751
+ use_cache = False
752
+
753
+ next_decoder_cache = () if use_cache else None
754
+ for i, layer_module in enumerate(self.layer):
755
+ if output_hidden_states:
756
+ all_hidden_states = all_hidden_states + (hidden_states,)
757
+
758
+ layer_head_mask = head_mask[i] if head_mask is not None else None
759
+ past_key_value = past_key_values[i] if past_key_values is not None else None
760
+
761
+ if self.gradient_checkpointing and self.training:
762
+
763
+ def create_custom_forward(module):
764
+ def custom_forward(*inputs):
765
+ return module(*inputs, past_key_value, output_attentions)
766
+
767
+ return custom_forward
768
+
769
+ layer_outputs = torch.utils.checkpoint.checkpoint(
770
+ create_custom_forward(layer_module),
771
+ hidden_states,
772
+ attention_mask,
773
+ layer_head_mask,
774
+ encoder_hidden_states,
775
+ encoder_attention_mask,
776
+ alibi_bias,
777
+ )
778
+ else:
779
+ layer_outputs = layer_module(
780
+ hidden_states,
781
+ attention_mask,
782
+ layer_head_mask,
783
+ encoder_hidden_states,
784
+ encoder_attention_mask,
785
+ alibi_bias,
786
+ past_key_value,
787
+ output_attentions,
788
+ )
789
+
790
+ hidden_states = layer_outputs[0]
791
+ if use_cache:
792
+ next_decoder_cache += (layer_outputs[-1],)
793
+ if output_attentions:
794
+ all_self_attentions = all_self_attentions + (layer_outputs[1],)
795
+ if self.config.add_cross_attention:
796
+ all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
797
+
798
+ if output_hidden_states:
799
+ all_hidden_states = all_hidden_states + (hidden_states,)
800
+
801
+ if not return_dict:
802
+ return tuple(
803
+ v
804
+ for v in [
805
+ hidden_states,
806
+ next_decoder_cache,
807
+ all_hidden_states,
808
+ all_self_attentions,
809
+ all_cross_attentions,
810
+ ]
811
+ if v is not None
812
+ )
813
+ return BaseModelOutputWithPastAndCrossAttentions(
814
+ last_hidden_state=hidden_states,
815
+ past_key_values=next_decoder_cache,
816
+ hidden_states=all_hidden_states,
817
+ attentions=all_self_attentions,
818
+ cross_attentions=all_cross_attentions,
819
+ )
820
+
821
+
822
+ class JinaBertPooler(nn.Module):
823
+ def __init__(self, config):
824
+ super().__init__()
825
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
826
+ self.activation = nn.Tanh()
827
+
828
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
829
+ # We "pool" the model by simply taking the hidden state corresponding
830
+ # to the first token.
831
+ first_token_tensor = hidden_states[:, 0]
832
+ pooled_output = self.dense(first_token_tensor)
833
+ pooled_output = self.activation(pooled_output)
834
+ return pooled_output
835
+
836
+
837
+ class JinaBertPredictionHeadTransform(nn.Module):
838
+ def __init__(self, config):
839
+ super().__init__()
840
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
841
+ if isinstance(config.hidden_act, str):
842
+ self.transform_act_fn = ACT2FN[config.hidden_act]
843
+ else:
844
+ self.transform_act_fn = config.hidden_act
845
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
846
+
847
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
848
+ hidden_states = self.dense(hidden_states)
849
+ hidden_states = self.transform_act_fn(hidden_states)
850
+ hidden_states = self.LayerNorm(hidden_states)
851
+ return hidden_states
852
+
853
+
854
+ class JinaBertLMPredictionHead(nn.Module):
855
+ def __init__(self, config):
856
+ super().__init__()
857
+ self.transform = JinaBertPredictionHeadTransform(config)
858
+
859
+ # The output weights are the same as the input embeddings, but there is
860
+ # an output-only bias for each token.
861
+ self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
862
+
863
+ self.bias = nn.Parameter(torch.zeros(config.vocab_size))
864
+
865
+ # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
866
+ self.decoder.bias = self.bias
867
+
868
+ def forward(self, hidden_states):
869
+ hidden_states = self.transform(hidden_states)
870
+ hidden_states = self.decoder(hidden_states)
871
+ return hidden_states
872
+
873
+
874
+ class JinaBertOnlyMLMHead(nn.Module):
875
+ def __init__(self, config):
876
+ super().__init__()
877
+ self.predictions = JinaBertLMPredictionHead(config)
878
+
879
+ def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
880
+ prediction_scores = self.predictions(sequence_output)
881
+ return prediction_scores
882
+
883
+
884
+ class JinaBertOnlyNSPHead(nn.Module):
885
+ def __init__(self, config):
886
+ super().__init__()
887
+ self.seq_relationship = nn.Linear(config.hidden_size, 2)
888
+
889
+ def forward(self, pooled_output):
890
+ seq_relationship_score = self.seq_relationship(pooled_output)
891
+ return seq_relationship_score
892
+
893
+
894
+ class JinaBertPreTrainingHeads(nn.Module):
895
+ def __init__(self, config):
896
+ super().__init__()
897
+ self.predictions = JinaBertLMPredictionHead(config)
898
+ self.seq_relationship = nn.Linear(config.hidden_size, 2)
899
+
900
+ def forward(self, sequence_output, pooled_output):
901
+ prediction_scores = self.predictions(sequence_output)
902
+ seq_relationship_score = self.seq_relationship(pooled_output)
903
+ return prediction_scores, seq_relationship_score
904
+
905
+
906
+ class JinaBertPreTrainedModel(PreTrainedModel):
907
+ """
908
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
909
+ models.
910
+ """
911
+
912
+ config_class = JinaBertConfig
913
+ load_tf_weights = load_tf_weights_in_bert
914
+ base_model_prefix = "bert"
915
+ supports_gradient_checkpointing = True
916
+ _no_split_modules = ["JinaBertLayer"]
917
+
918
+ def _init_weights(self, module):
919
+ """Initialize the weights"""
920
+ if isinstance(module, nn.Linear):
921
+ # Slightly different from the TF version which uses truncated_normal for initialization
922
+ # cf https://github.com/pytorch/pytorch/pull/5617
923
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
924
+ if module.bias is not None:
925
+ module.bias.data.zero_()
926
+ elif isinstance(module, nn.Embedding):
927
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
928
+ if module.padding_idx is not None:
929
+ module.weight.data[module.padding_idx].zero_()
930
+ elif isinstance(module, nn.LayerNorm):
931
+ module.bias.data.zero_()
932
+ module.weight.data.fill_(1.0)
933
+
934
+ def _set_gradient_checkpointing(self, module, value=False):
935
+ if isinstance(module, JinaBertEncoder):
936
+ module.gradient_checkpointing = value
937
+
938
+
939
+ @dataclass
940
+ class JinaBertForPreTrainingOutput(ModelOutput):
941
+ """
942
+ Output type of [`BertForPreTraining`].
943
+
944
+ Args:
945
+ loss (*optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
946
+ Total loss as the sum of the masked language modeling loss and the next sequence prediction
947
+ (classification) loss.
948
+ prediction_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
949
+ Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
950
+ seq_relationship_logits (`torch.FloatTensor` of shape `(batch_size, 2)`):
951
+ Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
952
+ before SoftMax).
953
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
954
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
955
+ shape `(batch_size, sequence_length, hidden_size)`.
956
+
957
+ Hidden-states of the model at the output of each layer plus the initial embedding outputs.
958
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
959
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
960
+ sequence_length)`.
961
+
962
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
963
+ heads.
964
+ """
965
+
966
+ loss: Optional[torch.FloatTensor] = None
967
+ prediction_logits: torch.FloatTensor = None
968
+ seq_relationship_logits: torch.FloatTensor = None
969
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
970
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
971
+
972
+
973
+ BERT_START_DOCSTRING = r"""
974
+
975
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
976
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
977
+ etc.)
978
+
979
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
980
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
981
+ and behavior.
982
+
983
+ Parameters:
984
+ config ([`BertConfig`]): Model configuration class with all the parameters of the model.
985
+ Initializing with a config file does not load the weights associated with the model, only the
986
+ configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
987
+ """
988
+
989
+ BERT_INPUTS_DOCSTRING = r"""
990
+ Args:
991
+ input_ids (`torch.LongTensor` of shape `({0})`):
992
+ Indices of input sequence tokens in the vocabulary.
993
+
994
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
995
+ [`PreTrainedTokenizer.__call__`] for details.
996
+
997
+ [What are input IDs?](../glossary#input-ids)
998
+ attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
999
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
1000
+
1001
+ - 1 for tokens that are **not masked**,
1002
+ - 0 for tokens that are **masked**.
1003
+
1004
+ [What are attention masks?](../glossary#attention-mask)
1005
+ token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
1006
+ Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
1007
+ 1]`:
1008
+
1009
+ - 0 corresponds to a *sentence A* token,
1010
+ - 1 corresponds to a *sentence B* token.
1011
+
1012
+ [What are token type IDs?](../glossary#token-type-ids)
1013
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
1014
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
1015
+ config.max_position_embeddings - 1]`.
1016
+
1017
+ [What are position IDs?](../glossary#position-ids)
1018
+ head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
1019
+ Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
1020
+
1021
+ - 1 indicates the head is **not masked**,
1022
+ - 0 indicates the head is **masked**.
1023
+
1024
+ inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
1025
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
1026
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
1027
+ model's internal embedding lookup matrix.
1028
+ output_attentions (`bool`, *optional*):
1029
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
1030
+ tensors for more detail.
1031
+ output_hidden_states (`bool`, *optional*):
1032
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
1033
+ more detail.
1034
+ return_dict (`bool`, *optional*):
1035
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
1036
+ """
1037
+
1038
+
1039
+ @add_start_docstrings(
1040
+ "The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
1041
+ BERT_START_DOCSTRING,
1042
+ )
1043
+ class JinaBertModel(JinaBertPreTrainedModel):
1044
+ """
1045
+
1046
+ The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
1047
+ cross-attention is added between the self-attention layers, following the architecture described in [Attention is
1048
+ all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
1049
+ Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
1050
+
1051
+ To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
1052
+ to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
1053
+ `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
1054
+ """
1055
+
1056
+ def __init__(self, config: JinaBertConfig, add_pooling_layer=True):
1057
+ super().__init__(config)
1058
+ self.config = config
1059
+
1060
+ self.emb_pooler = config.emb_pooler
1061
+ self._name_or_path = config._name_or_path
1062
+ if self.emb_pooler:
1063
+ from transformers import AutoTokenizer
1064
+
1065
+ self.tokenizer = AutoTokenizer.from_pretrained(config._name_or_path)
1066
+
1067
+ self.embeddings = JinaBertEmbeddings(config)
1068
+ self.encoder = JinaBertEncoder(config)
1069
+
1070
+ self.pooler = JinaBertPooler(config) if add_pooling_layer else None
1071
+
1072
+ # Initialize weights and apply final processing
1073
+ self.post_init()
1074
+
1075
+ @torch.inference_mode()
1076
+ def encode(
1077
+ self: 'JinaBertModel',
1078
+ sentences: Union[str, List[str]],
1079
+ batch_size: int = 32,
1080
+ show_progress_bar: Optional[bool] = None,
1081
+ output_value: str = 'sentence_embedding',
1082
+ convert_to_numpy: bool = True,
1083
+ convert_to_tensor: bool = False,
1084
+ device: Optional[torch.device] = None,
1085
+ normalize_embeddings: bool = False,
1086
+ **tokenizer_kwargs,
1087
+ ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
1088
+ """
1089
+ Computes sentence embeddings
1090
+
1091
+ Args:
1092
+ sentences(`str` or `List[str]`):
1093
+ Sentence or sentences to be encoded
1094
+ batch_size(`int`, *optional*, defaults to 32):
1095
+ Batch size for the computation
1096
+ show_progress_bar(`bool`, *optional*, defaults to None):
1097
+ Show a progress bar when encoding sentences.
1098
+ If set to None, progress bar is only shown when `logger.level == logging.INFO` or `logger.level == logging.DEBUG`.
1099
+ output_value(`str`, *optional*, defaults to 'sentence_embedding'):
1100
+ Default sentence_embedding, to get sentence embeddings.
1101
+ Can be set to token_embeddings to get wordpiece token embeddings.
1102
+ Set to None, to get all output values
1103
+ convert_to_numpy(`bool`, *optional*, defaults to True):
1104
+ If true, the output is a list of numpy vectors.
1105
+ Else, it is a list of pytorch tensors.
1106
+ convert_to_tensor(`bool`, *optional*, defaults to False):
1107
+ If true, you get one large tensor as return.
1108
+ Overwrites any setting from convert_to_numpy
1109
+ device(`torch.device`, *optional*, defaults to None):
1110
+ Which torch.device to use for the computation
1111
+ normalize_embeddings(`bool`, *optional*, defaults to False):
1112
+ If set to true, returned vectors will have length 1. In that case, the faster dot-product (util.dot_score) instead of cosine similarity can be used.
1113
+ tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
1114
+ Keyword arguments for the tokenizer
1115
+
1116
+ Returns:
1117
+ By default, a list of tensors is returned.
1118
+ If convert_to_tensor, a stacked tensor is returned.
1119
+ If convert_to_numpy, a numpy matrix is returned.
1120
+ """
1121
+ if not self.emb_pooler:
1122
+ warnings.warn("No emb_pooler specified, defaulting to mean pooling.")
1123
+ self.emb_pooler = 'mean'
1124
+ from transformers import AutoTokenizer
1125
+
1126
+ self.tokenizer = AutoTokenizer.from_pretrained(self._name_or_path)
1127
+ is_training = self.training
1128
+ self.eval()
1129
+
1130
+ if show_progress_bar is None:
1131
+ show_progress_bar = (
1132
+ logger.getEffectiveLevel() == logging.INFO
1133
+ or logger.getEffectiveLevel() == logging.DEBUG
1134
+ )
1135
+
1136
+ if convert_to_tensor:
1137
+ convert_to_numpy = False
1138
+
1139
+ if output_value != 'sentence_embedding':
1140
+ convert_to_tensor = False
1141
+ convert_to_numpy = False
1142
+
1143
+ input_was_string = False
1144
+ if isinstance(sentences, str) or not hasattr(sentences, '__len__'):
1145
+ sentences = [sentences]
1146
+ input_was_string = True
1147
+
1148
+ if device is not None:
1149
+ self.to(device)
1150
+
1151
+ # TODO: Maybe use better length heuristic?
1152
+ permutation = np.argsort([-len(i) for i in sentences])
1153
+ inverse_permutation = np.argsort(permutation)
1154
+ sentences = [sentences[idx] for idx in permutation]
1155
+
1156
+ tokenizer_kwargs['padding'] = tokenizer_kwargs.get('padding', True)
1157
+ tokenizer_kwargs['max_length'] = tokenizer_kwargs.get('max_length', 8192)
1158
+ tokenizer_kwargs['truncation'] = tokenizer_kwargs.get('truncation', True)
1159
+
1160
+ all_embeddings = []
1161
+
1162
+ if has_tqdm:
1163
+ range_iter = trange(
1164
+ 0,
1165
+ len(sentences),
1166
+ batch_size,
1167
+ desc="Encoding",
1168
+ disable=not show_progress_bar,
1169
+ )
1170
+ else:
1171
+ range_iter = range(0, len(sentences), batch_size)
1172
+
1173
+ for i in range_iter:
1174
+ encoded_input = self.tokenizer(
1175
+ sentences[i : i + batch_size],
1176
+ return_tensors='pt',
1177
+ **tokenizer_kwargs,
1178
+ ).to(self.device)
1179
+ token_embs = self.forward(**encoded_input)[0]
1180
+
1181
+ # Accumulate in fp32 to avoid overflow
1182
+ token_embs = token_embs.float()
1183
+
1184
+ if output_value == 'token_embeddings':
1185
+ raise NotImplementedError
1186
+ elif output_value is None:
1187
+ raise NotImplementedError
1188
+ else:
1189
+ embeddings = self.mean_pooling(
1190
+ token_embs, encoded_input['attention_mask']
1191
+ )
1192
+
1193
+ if normalize_embeddings:
1194
+ embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
1195
+
1196
+ if convert_to_numpy:
1197
+ embeddings = embeddings.cpu()
1198
+ all_embeddings.extend(embeddings)
1199
+
1200
+ all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
1201
+
1202
+ if convert_to_tensor:
1203
+ all_embeddings = torch.stack(all_embeddings)
1204
+ elif convert_to_numpy:
1205
+ all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])
1206
+
1207
+ if input_was_string:
1208
+ all_embeddings = all_embeddings[0]
1209
+
1210
+ self.train(is_training)
1211
+ return all_embeddings
1212
+
1213
+ def mean_pooling(
1214
+ self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
1215
+ ):
1216
+ input_mask_expanded = (
1217
+ attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
1218
+ )
1219
+ return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
1220
+ input_mask_expanded.sum(1), min=1e-9
1221
+ )
1222
+
1223
+ def get_input_embeddings(self):
1224
+ return self.embeddings.word_embeddings
1225
+
1226
+ def set_input_embeddings(self, value):
1227
+ self.embeddings.word_embeddings = value
1228
+
1229
+ def _prune_heads(self, heads_to_prune):
1230
+ """
1231
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
1232
+ class PreTrainedModel
1233
+ """
1234
+ for layer, heads in heads_to_prune.items():
1235
+ self.encoder.layer[layer].attention.prune_heads(heads)
1236
+
1237
+ @add_start_docstrings_to_model_forward(
1238
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1239
+ )
1240
+ @add_code_sample_docstrings(
1241
+ checkpoint=_CHECKPOINT_FOR_DOC,
1242
+ output_type=BaseModelOutputWithPoolingAndCrossAttentions,
1243
+ config_class=_CONFIG_FOR_DOC,
1244
+ )
1245
+ def forward(
1246
+ self,
1247
+ input_ids: Optional[torch.Tensor] = None,
1248
+ attention_mask: Optional[torch.Tensor] = None,
1249
+ token_type_ids: Optional[torch.Tensor] = None,
1250
+ position_ids: Optional[torch.Tensor] = None,
1251
+ head_mask: Optional[torch.Tensor] = None,
1252
+ inputs_embeds: Optional[torch.Tensor] = None,
1253
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1254
+ encoder_attention_mask: Optional[torch.Tensor] = None,
1255
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1256
+ use_cache: Optional[bool] = None,
1257
+ output_attentions: Optional[bool] = None,
1258
+ output_hidden_states: Optional[bool] = None,
1259
+ return_dict: Optional[bool] = None,
1260
+ ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
1261
+ r"""
1262
+ encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
1263
+ Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
1264
+ the model is configured as a decoder.
1265
+ encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
1266
+ Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
1267
+ the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
1268
+
1269
+ - 1 for tokens that are **not masked**,
1270
+ - 0 for tokens that are **masked**.
1271
+ past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
1272
+ Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
1273
+
1274
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
1275
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
1276
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
1277
+ use_cache (`bool`, *optional*):
1278
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
1279
+ `past_key_values`).
1280
+ """
1281
+ output_attentions = (
1282
+ output_attentions
1283
+ if output_attentions is not None
1284
+ else self.config.output_attentions
1285
+ )
1286
+ output_hidden_states = (
1287
+ output_hidden_states
1288
+ if output_hidden_states is not None
1289
+ else self.config.output_hidden_states
1290
+ )
1291
+ return_dict = (
1292
+ return_dict if return_dict is not None else self.config.use_return_dict
1293
+ )
1294
+
1295
+ if self.config.is_decoder:
1296
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
1297
+ else:
1298
+ use_cache = False
1299
+
1300
+ if input_ids is not None and inputs_embeds is not None:
1301
+ raise ValueError(
1302
+ "You cannot specify both input_ids and inputs_embeds at the same time"
1303
+ )
1304
+ elif input_ids is not None:
1305
+ # self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
1306
+ input_shape = input_ids.size()
1307
+ elif inputs_embeds is not None:
1308
+ input_shape = inputs_embeds.size()[:-1]
1309
+ else:
1310
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
1311
+
1312
+ batch_size, seq_length = input_shape
1313
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
1314
+
1315
+ # past_key_values_length
1316
+ past_key_values_length = (
1317
+ past_key_values[0][0].shape[2] if past_key_values is not None else 0
1318
+ )
1319
+
1320
+ if attention_mask is None:
1321
+ attention_mask = torch.ones(
1322
+ ((batch_size, seq_length + past_key_values_length)), device=device
1323
+ )
1324
+
1325
+ if token_type_ids is None:
1326
+ if hasattr(self.embeddings, "token_type_ids"):
1327
+ buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
1328
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
1329
+ batch_size, seq_length
1330
+ )
1331
+ token_type_ids = buffered_token_type_ids_expanded
1332
+ else:
1333
+ token_type_ids = torch.zeros(
1334
+ input_shape, dtype=torch.long, device=device
1335
+ )
1336
+
1337
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
1338
+ # ourselves in which case we just need to make it broadcastable to all heads.
1339
+ extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
1340
+ attention_mask, input_shape
1341
+ )
1342
+
1343
+ # If a 2D or 3D attention mask is provided for the cross-attention
1344
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
1345
+ if self.config.is_decoder and encoder_hidden_states is not None:
1346
+ (
1347
+ encoder_batch_size,
1348
+ encoder_sequence_length,
1349
+ _,
1350
+ ) = encoder_hidden_states.size()
1351
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
1352
+ if encoder_attention_mask is None:
1353
+ encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
1354
+ encoder_extended_attention_mask = self.invert_attention_mask(
1355
+ encoder_attention_mask
1356
+ )
1357
+ else:
1358
+ encoder_extended_attention_mask = None
1359
+
1360
+ # Prepare head mask if needed
1361
+ # 1.0 in head_mask indicate we keep the head
1362
+ # attention_probs has shape bsz x n_heads x N x N
1363
+ # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
1364
+ # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
1365
+ head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
1366
+
1367
+ embedding_output = self.embeddings(
1368
+ input_ids=input_ids,
1369
+ position_ids=position_ids,
1370
+ token_type_ids=token_type_ids,
1371
+ inputs_embeds=inputs_embeds,
1372
+ past_key_values_length=past_key_values_length,
1373
+ )
1374
+ encoder_outputs = self.encoder(
1375
+ embedding_output,
1376
+ attention_mask=extended_attention_mask,
1377
+ head_mask=head_mask,
1378
+ encoder_hidden_states=encoder_hidden_states,
1379
+ encoder_attention_mask=encoder_extended_attention_mask,
1380
+ past_key_values=past_key_values,
1381
+ use_cache=use_cache,
1382
+ output_attentions=output_attentions,
1383
+ output_hidden_states=output_hidden_states,
1384
+ return_dict=return_dict,
1385
+ )
1386
+ sequence_output = encoder_outputs[0]
1387
+ pooled_output = (
1388
+ self.pooler(sequence_output) if self.pooler is not None else None
1389
+ )
1390
+
1391
+ if not return_dict:
1392
+ return (sequence_output, pooled_output) + encoder_outputs[1:]
1393
+
1394
+ return BaseModelOutputWithPoolingAndCrossAttentions(
1395
+ last_hidden_state=sequence_output,
1396
+ pooler_output=pooled_output,
1397
+ past_key_values=encoder_outputs.past_key_values,
1398
+ hidden_states=encoder_outputs.hidden_states,
1399
+ attentions=encoder_outputs.attentions,
1400
+ cross_attentions=encoder_outputs.cross_attentions,
1401
+ )
1402
+
1403
+
1404
+ @add_start_docstrings(
1405
+ """
1406
+ Bert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next
1407
+ sentence prediction (classification)` head.
1408
+ """,
1409
+ BERT_START_DOCSTRING,
1410
+ )
1411
+ class JinaBertForPreTraining(JinaBertPreTrainedModel):
1412
+ _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
1413
+
1414
+ def __init__(self, config):
1415
+ super().__init__(config)
1416
+
1417
+ self.bert = JinaBertModel(config)
1418
+ self.cls = JinaBertPreTrainingHeads(config)
1419
+
1420
+ # Initialize weights and apply final processing
1421
+ self.post_init()
1422
+
1423
+ def get_output_embeddings(self):
1424
+ return self.cls.predictions.decoder
1425
+
1426
+ def set_output_embeddings(self, new_embeddings):
1427
+ self.cls.predictions.decoder = new_embeddings
1428
+
1429
+ @add_start_docstrings_to_model_forward(
1430
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1431
+ )
1432
+ @replace_return_docstrings(
1433
+ output_type=JinaBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC
1434
+ )
1435
+ def forward(
1436
+ self,
1437
+ input_ids: Optional[torch.Tensor] = None,
1438
+ attention_mask: Optional[torch.Tensor] = None,
1439
+ token_type_ids: Optional[torch.Tensor] = None,
1440
+ position_ids: Optional[torch.Tensor] = None,
1441
+ head_mask: Optional[torch.Tensor] = None,
1442
+ inputs_embeds: Optional[torch.Tensor] = None,
1443
+ labels: Optional[torch.Tensor] = None,
1444
+ next_sentence_label: Optional[torch.Tensor] = None,
1445
+ output_attentions: Optional[bool] = None,
1446
+ output_hidden_states: Optional[bool] = None,
1447
+ return_dict: Optional[bool] = None,
1448
+ ) -> Union[Tuple[torch.Tensor], JinaBertForPreTrainingOutput]:
1449
+ r"""
1450
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1451
+ Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
1452
+ config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked),
1453
+ the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
1454
+ next_sentence_label (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1455
+ Labels for computing the next sequence prediction (classification) loss. Input should be a sequence
1456
+ pair (see `input_ids` docstring) Indices should be in `[0, 1]`:
1457
+
1458
+ - 0 indicates sequence B is a continuation of sequence A,
1459
+ - 1 indicates sequence B is a random sequence.
1460
+ kwargs (`Dict[str, any]`, optional, defaults to *{}*):
1461
+ Used to hide legacy arguments that have been deprecated.
1462
+
1463
+ Returns:
1464
+ """
1465
+ return_dict = (
1466
+ return_dict if return_dict is not None else self.config.use_return_dict
1467
+ )
1468
+
1469
+ outputs = self.bert(
1470
+ input_ids,
1471
+ attention_mask=attention_mask,
1472
+ token_type_ids=token_type_ids,
1473
+ position_ids=position_ids,
1474
+ head_mask=head_mask,
1475
+ inputs_embeds=inputs_embeds,
1476
+ output_attentions=output_attentions,
1477
+ output_hidden_states=output_hidden_states,
1478
+ return_dict=return_dict,
1479
+ )
1480
+
1481
+ sequence_output, pooled_output = outputs[:2]
1482
+ prediction_scores, seq_relationship_score = self.cls(
1483
+ sequence_output, pooled_output
1484
+ )
1485
+
1486
+ total_loss = None
1487
+ if labels is not None and next_sentence_label is not None:
1488
+ loss_fct = CrossEntropyLoss()
1489
+ masked_lm_loss = loss_fct(
1490
+ prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
1491
+ )
1492
+ next_sentence_loss = loss_fct(
1493
+ seq_relationship_score.view(-1, 2), next_sentence_label.view(-1)
1494
+ )
1495
+ total_loss = masked_lm_loss + next_sentence_loss
1496
+
1497
+ if not return_dict:
1498
+ output = (prediction_scores, seq_relationship_score) + outputs[2:]
1499
+ return ((total_loss,) + output) if total_loss is not None else output
1500
+
1501
+ return JinaBertForPreTrainingOutput(
1502
+ loss=total_loss,
1503
+ prediction_logits=prediction_scores,
1504
+ seq_relationship_logits=seq_relationship_score,
1505
+ hidden_states=outputs.hidden_states,
1506
+ attentions=outputs.attentions,
1507
+ )
1508
+
1509
+
1510
+ @add_start_docstrings(
1511
+ """JinaBert Model with a `language modeling` head on top for CLM fine-tuning.""",
1512
+ BERT_START_DOCSTRING,
1513
+ )
1514
+ class JinaBertLMHeadModel(JinaBertPreTrainedModel):
1515
+ _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
1516
+
1517
+ def __init__(self, config):
1518
+ super().__init__(config)
1519
+
1520
+ if not config.is_decoder:
1521
+ logger.warning(
1522
+ "If you want to use `JinaBertLMHeadModel` as a standalone, add `is_decoder=True.`"
1523
+ )
1524
+
1525
+ self.bert = JinaBertModel(config, add_pooling_layer=False)
1526
+ self.cls = JinaBertOnlyMLMHead(config)
1527
+
1528
+ # Initialize weights and apply final processing
1529
+ self.post_init()
1530
+
1531
+ def get_output_embeddings(self):
1532
+ return self.cls.predictions.decoder
1533
+
1534
+ def set_output_embeddings(self, new_embeddings):
1535
+ self.cls.predictions.decoder = new_embeddings
1536
+
1537
+ @add_start_docstrings_to_model_forward(
1538
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1539
+ )
1540
+ @add_code_sample_docstrings(
1541
+ checkpoint=_CHECKPOINT_FOR_DOC,
1542
+ output_type=CausalLMOutputWithCrossAttentions,
1543
+ config_class=_CONFIG_FOR_DOC,
1544
+ )
1545
+ def forward(
1546
+ self,
1547
+ input_ids: Optional[torch.Tensor] = None,
1548
+ attention_mask: Optional[torch.Tensor] = None,
1549
+ token_type_ids: Optional[torch.Tensor] = None,
1550
+ position_ids: Optional[torch.Tensor] = None,
1551
+ head_mask: Optional[torch.Tensor] = None,
1552
+ inputs_embeds: Optional[torch.Tensor] = None,
1553
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1554
+ encoder_attention_mask: Optional[torch.Tensor] = None,
1555
+ labels: Optional[torch.Tensor] = None,
1556
+ past_key_values: Optional[List[torch.Tensor]] = None,
1557
+ use_cache: Optional[bool] = None,
1558
+ output_attentions: Optional[bool] = None,
1559
+ output_hidden_states: Optional[bool] = None,
1560
+ return_dict: Optional[bool] = None,
1561
+ ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
1562
+ r"""
1563
+ encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
1564
+ Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
1565
+ the model is configured as a decoder.
1566
+ encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
1567
+ Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
1568
+ the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
1569
+
1570
+ - 1 for tokens that are **not masked**,
1571
+ - 0 for tokens that are **masked**.
1572
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1573
+ Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
1574
+ `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
1575
+ ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`
1576
+ past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
1577
+ Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
1578
+
1579
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
1580
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
1581
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
1582
+ use_cache (`bool`, *optional*):
1583
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
1584
+ `past_key_values`).
1585
+ """
1586
+ return_dict = (
1587
+ return_dict if return_dict is not None else self.config.use_return_dict
1588
+ )
1589
+ if labels is not None:
1590
+ use_cache = False
1591
+
1592
+ outputs = self.bert(
1593
+ input_ids,
1594
+ attention_mask=attention_mask,
1595
+ token_type_ids=token_type_ids,
1596
+ position_ids=position_ids,
1597
+ head_mask=head_mask,
1598
+ inputs_embeds=inputs_embeds,
1599
+ encoder_hidden_states=encoder_hidden_states,
1600
+ encoder_attention_mask=encoder_attention_mask,
1601
+ past_key_values=past_key_values,
1602
+ use_cache=use_cache,
1603
+ output_attentions=output_attentions,
1604
+ output_hidden_states=output_hidden_states,
1605
+ return_dict=return_dict,
1606
+ )
1607
+
1608
+ sequence_output = outputs[0]
1609
+ prediction_scores = self.cls(sequence_output)
1610
+
1611
+ lm_loss = None
1612
+ if labels is not None:
1613
+ # we are doing next-token prediction; shift prediction scores and input ids by one
1614
+ shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
1615
+ labels = labels[:, 1:].contiguous()
1616
+ loss_fct = CrossEntropyLoss()
1617
+ lm_loss = loss_fct(
1618
+ shifted_prediction_scores.view(-1, self.config.vocab_size),
1619
+ labels.view(-1),
1620
+ )
1621
+
1622
+ if not return_dict:
1623
+ output = (prediction_scores,) + outputs[2:]
1624
+ return ((lm_loss,) + output) if lm_loss is not None else output
1625
+
1626
+ return CausalLMOutputWithCrossAttentions(
1627
+ loss=lm_loss,
1628
+ logits=prediction_scores,
1629
+ past_key_values=outputs.past_key_values,
1630
+ hidden_states=outputs.hidden_states,
1631
+ attentions=outputs.attentions,
1632
+ cross_attentions=outputs.cross_attentions,
1633
+ )
1634
+
1635
+ def prepare_inputs_for_generation(
1636
+ self,
1637
+ input_ids,
1638
+ past_key_values=None,
1639
+ attention_mask=None,
1640
+ use_cache=True,
1641
+ **model_kwargs,
1642
+ ):
1643
+ input_shape = input_ids.shape
1644
+ # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
1645
+ if attention_mask is None:
1646
+ attention_mask = input_ids.new_ones(input_shape)
1647
+
1648
+ # cut decoder_input_ids if past_key_values is used
1649
+ if past_key_values is not None:
1650
+ input_ids = input_ids[:, -1:]
1651
+
1652
+ return {
1653
+ "input_ids": input_ids,
1654
+ "attention_mask": attention_mask,
1655
+ "past_key_values": past_key_values,
1656
+ "use_cache": use_cache,
1657
+ }
1658
+
1659
+ def _reorder_cache(self, past_key_values, beam_idx):
1660
+ reordered_past = ()
1661
+ for layer_past in past_key_values:
1662
+ reordered_past += (
1663
+ tuple(
1664
+ past_state.index_select(0, beam_idx) for past_state in layer_past
1665
+ ),
1666
+ )
1667
+ return reordered_past
1668
+
1669
+
1670
+ @add_start_docstrings(
1671
+ """JinaBert Model with a `language modeling` head on top.""", BERT_START_DOCSTRING
1672
+ )
1673
+ class JinaBertForMaskedLM(JinaBertPreTrainedModel):
1674
+ _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
1675
+
1676
+ def __init__(self, config):
1677
+ super().__init__(config)
1678
+
1679
+ if config.is_decoder:
1680
+ logger.warning(
1681
+ "If you want to use `JinaBertForMaskedLM` make sure `config.is_decoder=False` for "
1682
+ "bi-directional self-attention."
1683
+ )
1684
+
1685
+ self.bert = JinaBertModel(config, add_pooling_layer=False)
1686
+ self.cls = JinaBertOnlyMLMHead(config)
1687
+
1688
+ # Initialize weights and apply final processing
1689
+ self.post_init()
1690
+
1691
+ def get_output_embeddings(self):
1692
+ return self.cls.predictions.decoder
1693
+
1694
+ def set_output_embeddings(self, new_embeddings):
1695
+ self.cls.predictions.decoder = new_embeddings
1696
+
1697
+ @add_start_docstrings_to_model_forward(
1698
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1699
+ )
1700
+ @add_code_sample_docstrings(
1701
+ checkpoint=_CHECKPOINT_FOR_DOC,
1702
+ output_type=MaskedLMOutput,
1703
+ config_class=_CONFIG_FOR_DOC,
1704
+ expected_output="'paris'",
1705
+ expected_loss=0.88,
1706
+ )
1707
+ def forward(
1708
+ self,
1709
+ input_ids: Optional[torch.Tensor] = None,
1710
+ attention_mask: Optional[torch.Tensor] = None,
1711
+ token_type_ids: Optional[torch.Tensor] = None,
1712
+ position_ids: Optional[torch.Tensor] = None,
1713
+ head_mask: Optional[torch.Tensor] = None,
1714
+ inputs_embeds: Optional[torch.Tensor] = None,
1715
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1716
+ encoder_attention_mask: Optional[torch.Tensor] = None,
1717
+ labels: Optional[torch.Tensor] = None,
1718
+ output_attentions: Optional[bool] = None,
1719
+ output_hidden_states: Optional[bool] = None,
1720
+ return_dict: Optional[bool] = None,
1721
+ ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
1722
+ r"""
1723
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1724
+ Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
1725
+ config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
1726
+ loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
1727
+ """
1728
+
1729
+ return_dict = (
1730
+ return_dict if return_dict is not None else self.config.use_return_dict
1731
+ )
1732
+
1733
+ outputs = self.bert(
1734
+ input_ids,
1735
+ attention_mask=attention_mask,
1736
+ token_type_ids=token_type_ids,
1737
+ position_ids=position_ids,
1738
+ head_mask=head_mask,
1739
+ inputs_embeds=inputs_embeds,
1740
+ encoder_hidden_states=encoder_hidden_states,
1741
+ encoder_attention_mask=encoder_attention_mask,
1742
+ output_attentions=output_attentions,
1743
+ output_hidden_states=output_hidden_states,
1744
+ return_dict=return_dict,
1745
+ )
1746
+
1747
+ sequence_output = outputs[0]
1748
+ prediction_scores = self.cls(sequence_output)
1749
+
1750
+ masked_lm_loss = None
1751
+ if labels is not None:
1752
+ loss_fct = CrossEntropyLoss() # -100 index = padding token
1753
+ masked_lm_loss = loss_fct(
1754
+ prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
1755
+ )
1756
+
1757
+ if not return_dict:
1758
+ output = (prediction_scores,) + outputs[2:]
1759
+ return (
1760
+ ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
1761
+ )
1762
+
1763
+ return MaskedLMOutput(
1764
+ loss=masked_lm_loss,
1765
+ logits=prediction_scores,
1766
+ hidden_states=outputs.hidden_states,
1767
+ attentions=outputs.attentions,
1768
+ )
1769
+
1770
+ def prepare_inputs_for_generation(
1771
+ self, input_ids, attention_mask=None, **model_kwargs
1772
+ ):
1773
+ input_shape = input_ids.shape
1774
+ effective_batch_size = input_shape[0]
1775
+
1776
+ # add a dummy token
1777
+ if self.config.pad_token_id is None:
1778
+ raise ValueError("The PAD token should be defined for generation")
1779
+
1780
+ attention_mask = torch.cat(
1781
+ [attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))],
1782
+ dim=-1,
1783
+ )
1784
+ dummy_token = torch.full(
1785
+ (effective_batch_size, 1),
1786
+ self.config.pad_token_id,
1787
+ dtype=torch.long,
1788
+ device=input_ids.device,
1789
+ )
1790
+ input_ids = torch.cat([input_ids, dummy_token], dim=1)
1791
+
1792
+ return {"input_ids": input_ids, "attention_mask": attention_mask}
1793
+
1794
+
1795
+ @add_start_docstrings(
1796
+ """JinaBert Model with a `next sentence prediction (classification)` head on top.""",
1797
+ BERT_START_DOCSTRING,
1798
+ )
1799
+ class JinaBertForNextSentencePrediction(JinaBertPreTrainedModel):
1800
+ def __init__(self, config):
1801
+ super().__init__(config)
1802
+
1803
+ self.bert = JinaBertModel(config)
1804
+ self.cls = JinaBertOnlyNSPHead(config)
1805
+
1806
+ # Initialize weights and apply final processing
1807
+ self.post_init()
1808
+
1809
+ @add_start_docstrings_to_model_forward(
1810
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1811
+ )
1812
+ @replace_return_docstrings(
1813
+ output_type=NextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC
1814
+ )
1815
+ def forward(
1816
+ self,
1817
+ input_ids: Optional[torch.Tensor] = None,
1818
+ attention_mask: Optional[torch.Tensor] = None,
1819
+ token_type_ids: Optional[torch.Tensor] = None,
1820
+ position_ids: Optional[torch.Tensor] = None,
1821
+ head_mask: Optional[torch.Tensor] = None,
1822
+ inputs_embeds: Optional[torch.Tensor] = None,
1823
+ labels: Optional[torch.Tensor] = None,
1824
+ output_attentions: Optional[bool] = None,
1825
+ output_hidden_states: Optional[bool] = None,
1826
+ return_dict: Optional[bool] = None,
1827
+ **kwargs,
1828
+ ) -> Union[Tuple[torch.Tensor], NextSentencePredictorOutput]:
1829
+ r"""
1830
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1831
+ Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
1832
+ (see `input_ids` docstring). Indices should be in `[0, 1]`:
1833
+
1834
+ - 0 indicates sequence B is a continuation of sequence A,
1835
+ - 1 indicates sequence B is a random sequence.
1836
+
1837
+ Returns:
1838
+ """
1839
+
1840
+ if "next_sentence_label" in kwargs:
1841
+ warnings.warn(
1842
+ "The `next_sentence_label` argument is deprecated and will be removed in a future version, use"
1843
+ " `labels` instead.",
1844
+ FutureWarning,
1845
+ )
1846
+ labels = kwargs.pop("next_sentence_label")
1847
+
1848
+ return_dict = (
1849
+ return_dict if return_dict is not None else self.config.use_return_dict
1850
+ )
1851
+
1852
+ outputs = self.bert(
1853
+ input_ids,
1854
+ attention_mask=attention_mask,
1855
+ token_type_ids=token_type_ids,
1856
+ position_ids=position_ids,
1857
+ head_mask=head_mask,
1858
+ inputs_embeds=inputs_embeds,
1859
+ output_attentions=output_attentions,
1860
+ output_hidden_states=output_hidden_states,
1861
+ return_dict=return_dict,
1862
+ )
1863
+
1864
+ pooled_output = outputs[1]
1865
+
1866
+ seq_relationship_scores = self.cls(pooled_output)
1867
+
1868
+ next_sentence_loss = None
1869
+ if labels is not None:
1870
+ loss_fct = CrossEntropyLoss()
1871
+ next_sentence_loss = loss_fct(
1872
+ seq_relationship_scores.view(-1, 2), labels.view(-1)
1873
+ )
1874
+
1875
+ if not return_dict:
1876
+ output = (seq_relationship_scores,) + outputs[2:]
1877
+ return (
1878
+ ((next_sentence_loss,) + output)
1879
+ if next_sentence_loss is not None
1880
+ else output
1881
+ )
1882
+
1883
+ return NextSentencePredictorOutput(
1884
+ loss=next_sentence_loss,
1885
+ logits=seq_relationship_scores,
1886
+ hidden_states=outputs.hidden_states,
1887
+ attentions=outputs.attentions,
1888
+ )
1889
+
1890
+
1891
+ @add_start_docstrings(
1892
+ """
1893
+ JinaBert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
1894
+ output) e.g. for GLUE tasks.
1895
+ """,
1896
+ BERT_START_DOCSTRING,
1897
+ )
1898
+ class JinaBertForSequenceClassification(JinaBertPreTrainedModel):
1899
+ def __init__(self, config):
1900
+ super().__init__(config)
1901
+ self.num_labels = config.num_labels
1902
+ self.config = config
1903
+
1904
+ self.bert = JinaBertModel(config)
1905
+ classifier_dropout = (
1906
+ config.classifier_dropout
1907
+ if config.classifier_dropout is not None
1908
+ else config.hidden_dropout_prob
1909
+ )
1910
+ self.dropout = nn.Dropout(classifier_dropout)
1911
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1912
+
1913
+ # Initialize weights and apply final processing
1914
+ self.post_init()
1915
+
1916
+ @add_start_docstrings_to_model_forward(
1917
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
1918
+ )
1919
+ @add_code_sample_docstrings(
1920
+ checkpoint=_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION,
1921
+ output_type=SequenceClassifierOutput,
1922
+ config_class=_CONFIG_FOR_DOC,
1923
+ expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
1924
+ expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
1925
+ )
1926
+ def forward(
1927
+ self,
1928
+ input_ids: Optional[torch.Tensor] = None,
1929
+ attention_mask: Optional[torch.Tensor] = None,
1930
+ token_type_ids: Optional[torch.Tensor] = None,
1931
+ position_ids: Optional[torch.Tensor] = None,
1932
+ head_mask: Optional[torch.Tensor] = None,
1933
+ inputs_embeds: Optional[torch.Tensor] = None,
1934
+ labels: Optional[torch.Tensor] = None,
1935
+ output_attentions: Optional[bool] = None,
1936
+ output_hidden_states: Optional[bool] = None,
1937
+ return_dict: Optional[bool] = None,
1938
+ ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
1939
+ r"""
1940
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1941
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1942
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1943
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1944
+ """
1945
+ return_dict = (
1946
+ return_dict if return_dict is not None else self.config.use_return_dict
1947
+ )
1948
+
1949
+ outputs = self.bert(
1950
+ input_ids,
1951
+ attention_mask=attention_mask,
1952
+ token_type_ids=token_type_ids,
1953
+ position_ids=position_ids,
1954
+ head_mask=head_mask,
1955
+ inputs_embeds=inputs_embeds,
1956
+ output_attentions=output_attentions,
1957
+ output_hidden_states=output_hidden_states,
1958
+ return_dict=return_dict,
1959
+ )
1960
+
1961
+ pooled_output = outputs[1]
1962
+
1963
+ pooled_output = self.dropout(pooled_output)
1964
+ logits = self.classifier(pooled_output)
1965
+
1966
+ loss = None
1967
+ if labels is not None:
1968
+ if self.config.problem_type is None:
1969
+ if self.num_labels == 1:
1970
+ self.config.problem_type = "regression"
1971
+ elif self.num_labels > 1 and (
1972
+ labels.dtype == torch.long or labels.dtype == torch.int
1973
+ ):
1974
+ self.config.problem_type = "single_label_classification"
1975
+ else:
1976
+ self.config.problem_type = "multi_label_classification"
1977
+
1978
+ if self.config.problem_type == "regression":
1979
+ loss_fct = MSELoss()
1980
+ if self.num_labels == 1:
1981
+ loss = loss_fct(logits.squeeze(), labels.squeeze())
1982
+ else:
1983
+ loss = loss_fct(logits, labels)
1984
+ elif self.config.problem_type == "single_label_classification":
1985
+ loss_fct = CrossEntropyLoss()
1986
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1987
+ elif self.config.problem_type == "multi_label_classification":
1988
+ loss_fct = BCEWithLogitsLoss()
1989
+ loss = loss_fct(logits, labels)
1990
+ if not return_dict:
1991
+ output = (logits,) + outputs[2:]
1992
+ return ((loss,) + output) if loss is not None else output
1993
+
1994
+ return SequenceClassifierOutput(
1995
+ loss=loss,
1996
+ logits=logits,
1997
+ hidden_states=outputs.hidden_states,
1998
+ attentions=outputs.attentions,
1999
+ )
2000
+
2001
+
2002
+ @add_start_docstrings(
2003
+ """
2004
+ JinaBert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
2005
+ softmax) e.g. for RocStories/SWAG tasks.
2006
+ """,
2007
+ BERT_START_DOCSTRING,
2008
+ )
2009
+ class JinaBertForMultipleChoice(JinaBertPreTrainedModel):
2010
+ def __init__(self, config):
2011
+ super().__init__(config)
2012
+
2013
+ self.bert = JinaBertModel(config)
2014
+ classifier_dropout = (
2015
+ config.classifier_dropout
2016
+ if config.classifier_dropout is not None
2017
+ else config.hidden_dropout_prob
2018
+ )
2019
+ self.dropout = nn.Dropout(classifier_dropout)
2020
+ self.classifier = nn.Linear(config.hidden_size, 1)
2021
+
2022
+ # Initialize weights and apply final processing
2023
+ self.post_init()
2024
+
2025
+ @add_start_docstrings_to_model_forward(
2026
+ BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
2027
+ )
2028
+ @add_code_sample_docstrings(
2029
+ checkpoint=_CHECKPOINT_FOR_DOC,
2030
+ output_type=MultipleChoiceModelOutput,
2031
+ config_class=_CONFIG_FOR_DOC,
2032
+ )
2033
+ def forward(
2034
+ self,
2035
+ input_ids: Optional[torch.Tensor] = None,
2036
+ attention_mask: Optional[torch.Tensor] = None,
2037
+ token_type_ids: Optional[torch.Tensor] = None,
2038
+ position_ids: Optional[torch.Tensor] = None,
2039
+ head_mask: Optional[torch.Tensor] = None,
2040
+ inputs_embeds: Optional[torch.Tensor] = None,
2041
+ labels: Optional[torch.Tensor] = None,
2042
+ output_attentions: Optional[bool] = None,
2043
+ output_hidden_states: Optional[bool] = None,
2044
+ return_dict: Optional[bool] = None,
2045
+ ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
2046
+ r"""
2047
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
2048
+ Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
2049
+ num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
2050
+ `input_ids` above)
2051
+ """
2052
+ return_dict = (
2053
+ return_dict if return_dict is not None else self.config.use_return_dict
2054
+ )
2055
+ num_choices = (
2056
+ input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
2057
+ )
2058
+
2059
+ input_ids = (
2060
+ input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
2061
+ )
2062
+ attention_mask = (
2063
+ attention_mask.view(-1, attention_mask.size(-1))
2064
+ if attention_mask is not None
2065
+ else None
2066
+ )
2067
+ token_type_ids = (
2068
+ token_type_ids.view(-1, token_type_ids.size(-1))
2069
+ if token_type_ids is not None
2070
+ else None
2071
+ )
2072
+ position_ids = (
2073
+ position_ids.view(-1, position_ids.size(-1))
2074
+ if position_ids is not None
2075
+ else None
2076
+ )
2077
+ inputs_embeds = (
2078
+ inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
2079
+ if inputs_embeds is not None
2080
+ else None
2081
+ )
2082
+
2083
+ outputs = self.bert(
2084
+ input_ids,
2085
+ attention_mask=attention_mask,
2086
+ token_type_ids=token_type_ids,
2087
+ position_ids=position_ids,
2088
+ head_mask=head_mask,
2089
+ inputs_embeds=inputs_embeds,
2090
+ output_attentions=output_attentions,
2091
+ output_hidden_states=output_hidden_states,
2092
+ return_dict=return_dict,
2093
+ )
2094
+
2095
+ pooled_output = outputs[1]
2096
+
2097
+ pooled_output = self.dropout(pooled_output)
2098
+ logits = self.classifier(pooled_output)
2099
+ reshaped_logits = logits.view(-1, num_choices)
2100
+
2101
+ loss = None
2102
+ if labels is not None:
2103
+ loss_fct = CrossEntropyLoss()
2104
+ loss = loss_fct(reshaped_logits, labels)
2105
+
2106
+ if not return_dict:
2107
+ output = (reshaped_logits,) + outputs[2:]
2108
+ return ((loss,) + output) if loss is not None else output
2109
+
2110
+ return MultipleChoiceModelOutput(
2111
+ loss=loss,
2112
+ logits=reshaped_logits,
2113
+ hidden_states=outputs.hidden_states,
2114
+ attentions=outputs.attentions,
2115
+ )
2116
+
2117
+
2118
+ @add_start_docstrings(
2119
+ """
2120
+ JinaBert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
2121
+ Named-Entity-Recognition (NER) tasks.
2122
+ """,
2123
+ BERT_START_DOCSTRING,
2124
+ )
2125
+ class JinaBertForTokenClassification(JinaBertPreTrainedModel):
2126
+ def __init__(self, config):
2127
+ super().__init__(config)
2128
+ self.num_labels = config.num_labels
2129
+
2130
+ self.bert = JinaBertModel(config, add_pooling_layer=False)
2131
+ classifier_dropout = (
2132
+ config.classifier_dropout
2133
+ if config.classifier_dropout is not None
2134
+ else config.hidden_dropout_prob
2135
+ )
2136
+ self.dropout = nn.Dropout(classifier_dropout)
2137
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
2138
+
2139
+ # Initialize weights and apply final processing
2140
+ self.post_init()
2141
+
2142
+ @add_start_docstrings_to_model_forward(
2143
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
2144
+ )
2145
+ @add_code_sample_docstrings(
2146
+ checkpoint=_CHECKPOINT_FOR_TOKEN_CLASSIFICATION,
2147
+ output_type=TokenClassifierOutput,
2148
+ config_class=_CONFIG_FOR_DOC,
2149
+ expected_output=_TOKEN_CLASS_EXPECTED_OUTPUT,
2150
+ expected_loss=_TOKEN_CLASS_EXPECTED_LOSS,
2151
+ )
2152
+ def forward(
2153
+ self,
2154
+ input_ids: Optional[torch.Tensor] = None,
2155
+ attention_mask: Optional[torch.Tensor] = None,
2156
+ token_type_ids: Optional[torch.Tensor] = None,
2157
+ position_ids: Optional[torch.Tensor] = None,
2158
+ head_mask: Optional[torch.Tensor] = None,
2159
+ inputs_embeds: Optional[torch.Tensor] = None,
2160
+ labels: Optional[torch.Tensor] = None,
2161
+ output_attentions: Optional[bool] = None,
2162
+ output_hidden_states: Optional[bool] = None,
2163
+ return_dict: Optional[bool] = None,
2164
+ ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
2165
+ r"""
2166
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
2167
+ Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
2168
+ """
2169
+ return_dict = (
2170
+ return_dict if return_dict is not None else self.config.use_return_dict
2171
+ )
2172
+
2173
+ outputs = self.bert(
2174
+ input_ids,
2175
+ attention_mask=attention_mask,
2176
+ token_type_ids=token_type_ids,
2177
+ position_ids=position_ids,
2178
+ head_mask=head_mask,
2179
+ inputs_embeds=inputs_embeds,
2180
+ output_attentions=output_attentions,
2181
+ output_hidden_states=output_hidden_states,
2182
+ return_dict=return_dict,
2183
+ )
2184
+
2185
+ sequence_output = outputs[0]
2186
+
2187
+ sequence_output = self.dropout(sequence_output)
2188
+ logits = self.classifier(sequence_output)
2189
+
2190
+ loss = None
2191
+ if labels is not None:
2192
+ loss_fct = CrossEntropyLoss()
2193
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
2194
+
2195
+ if not return_dict:
2196
+ output = (logits,) + outputs[2:]
2197
+ return ((loss,) + output) if loss is not None else output
2198
+
2199
+ return TokenClassifierOutput(
2200
+ loss=loss,
2201
+ logits=logits,
2202
+ hidden_states=outputs.hidden_states,
2203
+ attentions=outputs.attentions,
2204
+ )
2205
+
2206
+
2207
+ @add_start_docstrings(
2208
+ """
2209
+ JinaBert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
2210
+ layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
2211
+ """,
2212
+ BERT_START_DOCSTRING,
2213
+ )
2214
+ class JinaBertForQuestionAnswering(JinaBertPreTrainedModel):
2215
+ def __init__(self, config):
2216
+ super().__init__(config)
2217
+ self.num_labels = config.num_labels
2218
+
2219
+ self.bert = JinaBertModel(config, add_pooling_layer=False)
2220
+ self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
2221
+
2222
+ # Initialize weights and apply final processing
2223
+ self.post_init()
2224
+
2225
+ @add_start_docstrings_to_model_forward(
2226
+ BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")
2227
+ )
2228
+ @add_code_sample_docstrings(
2229
+ checkpoint=_CHECKPOINT_FOR_QA,
2230
+ output_type=QuestionAnsweringModelOutput,
2231
+ config_class=_CONFIG_FOR_DOC,
2232
+ qa_target_start_index=_QA_TARGET_START_INDEX,
2233
+ qa_target_end_index=_QA_TARGET_END_INDEX,
2234
+ expected_output=_QA_EXPECTED_OUTPUT,
2235
+ expected_loss=_QA_EXPECTED_LOSS,
2236
+ )
2237
+ def forward(
2238
+ self,
2239
+ input_ids: Optional[torch.Tensor] = None,
2240
+ attention_mask: Optional[torch.Tensor] = None,
2241
+ token_type_ids: Optional[torch.Tensor] = None,
2242
+ position_ids: Optional[torch.Tensor] = None,
2243
+ head_mask: Optional[torch.Tensor] = None,
2244
+ inputs_embeds: Optional[torch.Tensor] = None,
2245
+ start_positions: Optional[torch.Tensor] = None,
2246
+ end_positions: Optional[torch.Tensor] = None,
2247
+ output_attentions: Optional[bool] = None,
2248
+ output_hidden_states: Optional[bool] = None,
2249
+ return_dict: Optional[bool] = None,
2250
+ ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
2251
+ r"""
2252
+ start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
2253
+ Labels for position (index) of the start of the labelled span for computing the token classification loss.
2254
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
2255
+ are not taken into account for computing the loss.
2256
+ end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
2257
+ Labels for position (index) of the end of the labelled span for computing the token classification loss.
2258
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
2259
+ are not taken into account for computing the loss.
2260
+ """
2261
+ return_dict = (
2262
+ return_dict if return_dict is not None else self.config.use_return_dict
2263
+ )
2264
+
2265
+ outputs = self.bert(
2266
+ input_ids,
2267
+ attention_mask=attention_mask,
2268
+ token_type_ids=token_type_ids,
2269
+ position_ids=position_ids,
2270
+ head_mask=head_mask,
2271
+ inputs_embeds=inputs_embeds,
2272
+ output_attentions=output_attentions,
2273
+ output_hidden_states=output_hidden_states,
2274
+ return_dict=return_dict,
2275
+ )
2276
+
2277
+ sequence_output = outputs[0]
2278
+
2279
+ logits = self.qa_outputs(sequence_output)
2280
+ start_logits, end_logits = logits.split(1, dim=-1)
2281
+ start_logits = start_logits.squeeze(-1).contiguous()
2282
+ end_logits = end_logits.squeeze(-1).contiguous()
2283
+
2284
+ total_loss = None
2285
+ if start_positions is not None and end_positions is not None:
2286
+ # If we are on multi-GPU, split add a dimension
2287
+ if len(start_positions.size()) > 1:
2288
+ start_positions = start_positions.squeeze(-1)
2289
+ if len(end_positions.size()) > 1:
2290
+ end_positions = end_positions.squeeze(-1)
2291
+ # sometimes the start/end positions are outside our model inputs, we ignore these terms
2292
+ ignored_index = start_logits.size(1)
2293
+ start_positions = start_positions.clamp(0, ignored_index)
2294
+ end_positions = end_positions.clamp(0, ignored_index)
2295
+
2296
+ loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
2297
+ start_loss = loss_fct(start_logits, start_positions)
2298
+ end_loss = loss_fct(end_logits, end_positions)
2299
+ total_loss = (start_loss + end_loss) / 2
2300
+
2301
+ if not return_dict:
2302
+ output = (start_logits, end_logits) + outputs[2:]
2303
+ return ((total_loss,) + output) if total_loss is not None else output
2304
+
2305
+ return QuestionAnsweringModelOutput(
2306
+ loss=total_loss,
2307
+ start_logits=start_logits,
2308
+ end_logits=end_logits,
2309
+ hidden_states=outputs.hidden_states,
2310
+ attentions=outputs.attentions,
2311
+ )
2312
+
modules.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ {
15
+ "idx": 2,
16
+ "name": "2",
17
+ "path": "2_Normalize",
18
+ "type": "sentence_transformers.models.Normalize"
19
+ }
20
+ ]
sentence_bert_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "max_seq_length": 8192,
3
+ "do_lower_case": false
4
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "cls_token": {
10
+ "content": "<s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "eos_token": {
17
+ "content": "</s>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "mask_token": {
24
+ "content": "<mask>",
25
+ "lstrip": true,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ },
30
+ "pad_token": {
31
+ "content": "<pad>",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false
36
+ },
37
+ "sep_token": {
38
+ "content": "</s>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": false,
42
+ "single_word": false
43
+ },
44
+ "unk_token": {
45
+ "content": "<unk>",
46
+ "lstrip": false,
47
+ "normalized": false,
48
+ "rstrip": false,
49
+ "single_word": false
50
+ }
51
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "added_tokens_decoder": {
4
+ "0": {
5
+ "content": "<s>",
6
+ "lstrip": false,
7
+ "normalized": false,
8
+ "rstrip": false,
9
+ "single_word": false,
10
+ "special": true
11
+ },
12
+ "1": {
13
+ "content": "<pad>",
14
+ "lstrip": false,
15
+ "normalized": false,
16
+ "rstrip": false,
17
+ "single_word": false,
18
+ "special": true
19
+ },
20
+ "2": {
21
+ "content": "</s>",
22
+ "lstrip": false,
23
+ "normalized": false,
24
+ "rstrip": false,
25
+ "single_word": false,
26
+ "special": true
27
+ },
28
+ "3": {
29
+ "content": "<unk>",
30
+ "lstrip": false,
31
+ "normalized": false,
32
+ "rstrip": false,
33
+ "single_word": false,
34
+ "special": true
35
+ },
36
+ "4": {
37
+ "content": "<mask>",
38
+ "lstrip": true,
39
+ "normalized": false,
40
+ "rstrip": false,
41
+ "single_word": false,
42
+ "special": true
43
+ }
44
+ },
45
+ "bos_token": "<s>",
46
+ "clean_up_tokenization_spaces": true,
47
+ "cls_token": "<s>",
48
+ "eos_token": "</s>",
49
+ "errors": "replace",
50
+ "mask_token": "<mask>",
51
+ "model_max_length": 8192,
52
+ "pad_token": "<pad>",
53
+ "sep_token": "</s>",
54
+ "tokenizer_class": "RobertaTokenizer",
55
+ "trim_offsets": true,
56
+ "unk_token": "<unk>"
57
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b0ee26af8356212d8f7cc241614b17132bd3315fb7287cad8191d18b0fdd1a0
3
+ size 5176
vocab.json ADDED
The diff for this file is too large to render. See raw diff