1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109
// Copyright 2019-present, the HuggingFace Inc. team, The Google AI Language Team and Facebook, Inc. // Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. // Copyright 2019 Guillaume Becquin // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // http://www.apache.org/licenses/LICENSE-2.0 // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. use crate::bert::embeddings::{BertEmbedding, BertEmbeddings}; use crate::bert::encoder::{BertEncoder, BertPooler}; use crate::common::activations::{_gelu, _mish, _relu}; use crate::common::dropout::Dropout; use crate::common::linear::{linear_no_bias, LinearNoBias}; use crate::Config; use serde::{Deserialize, Serialize}; use std::borrow::Borrow; use std::collections::HashMap; use tch::kind::Kind::Float; use tch::nn::Init; use tch::{nn, Kind, Tensor}; /// # BERT Pretrained model weight files pub struct BertModelResources; /// # BERT Pretrained model config files pub struct BertConfigResources; /// # BERT Pretrained model vocab files pub struct BertVocabResources; impl BertModelResources { /// Shared under Apache 2.0 license by the Google team at https://github.com/google-research/bert. Modified with conversion to C-array format. pub const BERT: (&'static str, &'static str) = ( "bert/model.ot", "https://cdn.huggingface.co/bert-base-uncased-rust_model.ot", ); /// Shared under MIT license by the MDZ Digital Library team at the Bavarian State Library at https://github.com/dbmdz/berts. Modified with conversion to C-array format. pub const BERT_NER: (&'static str, &'static str) = ( "bert-ner/model.ot", "https://cdn.huggingface.co/dbmdz/bert-large-cased-finetuned-conll03-english/rust_model.ot", ); /// Shared under Apache 2.0 license by Hugging Face Inc at https://github.com/huggingface/transformers/tree/master/examples/question-answering. Modified with conversion to C-array format. pub const BERT_QA: (&'static str, &'static str) = ( "bert-qa/model.ot", "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad-rust_model.ot", ); } impl BertConfigResources { /// Shared under Apache 2.0 license by the Google team at https://github.com/google-research/bert. Modified with conversion to C-array format. pub const BERT: (&'static str, &'static str) = ( "bert/config.json", "https://cdn.huggingface.co/bert-base-uncased-config.json", ); /// Shared under MIT license by the MDZ Digital Library team at the Bavarian State Library at https://github.com/dbmdz/berts. Modified with conversion to C-array format. pub const BERT_NER: (&'static str, &'static str) = ( "bert-ner/config.json", "https://cdn.huggingface.co/dbmdz/bert-large-cased-finetuned-conll03-english/config.json", ); /// Shared under Apache 2.0 license by Hugging Face Inc at https://github.com/huggingface/transformers/tree/master/examples/question-answering. Modified with conversion to C-array format. pub const BERT_QA: (&'static str, &'static str) = ( "bert-qa/config.json", "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json", ); } impl BertVocabResources { /// Shared under Apache 2.0 license by the Google team at https://github.com/google-research/bert. Modified with conversion to C-array format. pub const BERT: (&'static str, &'static str) = ( "bert/vocab.txt", "https://cdn.huggingface.co/bert-base-uncased-vocab.txt", ); /// Shared under MIT license by the MDZ Digital Library team at the Bavarian State Library at https://github.com/dbmdz/berts. Modified with conversion to C-array format. pub const BERT_NER: (&'static str, &'static str) = ( "bert-ner/vocab.txt", "https://cdn.huggingface.co/dbmdz/bert-large-cased-finetuned-conll03-english/vocab.txt", ); /// Shared under Apache 2.0 license by Hugging Face Inc at https://github.com/huggingface/transformers/tree/master/examples/question-answering. Modified with conversion to C-array format. pub const BERT_QA: (&'static str, &'static str) = ( "bert-qa/vocab.txt", "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad-vocab.txt", ); } #[allow(non_camel_case_types)] #[derive(Clone, Debug, Serialize, Deserialize)] /// # Activation function used in the attention layer and masked language model head pub enum Activation { /// Gaussian Error Linear Unit ([Hendrycks et al., 2016,](https://arxiv.org/abs/1606.08415)) gelu, /// Rectified Linear Unit relu, /// Mish ([Misra, 2019](https://arxiv.org/abs/1908.08681)) mish, } #[derive(Debug, Serialize, Deserialize)] /// # BERT model configuration /// Defines the BERT model architecture (e.g. number of layers, hidden layer size, label mapping...) pub struct BertConfig { pub hidden_act: Activation, pub attention_probs_dropout_prob: f64, pub hidden_dropout_prob: f64, pub hidden_size: i64, pub initializer_range: f32, pub intermediate_size: i64, pub max_position_embeddings: i64, pub num_attention_heads: i64, pub num_hidden_layers: i64, pub type_vocab_size: i64, pub vocab_size: i64, pub output_attentions: Option<bool>, pub output_hidden_states: Option<bool>, pub is_decoder: Option<bool>, pub id2label: Option<HashMap<i64, String>>, pub label2id: Option<HashMap<String, i64>>, } impl Config<BertConfig> for BertConfig {} /// # BERT Base model /// Base architecture for BERT models. Task-specific models will be built from this common base model /// It is made of the following blocks: /// - `embeddings`: `token`, `position` and `segment_id` embeddings /// - `encoder`: Encoder (transformer) made of a vector of layers. Each layer is made of a self-attention layer, an intermediate (linear) and output (linear + layer norm) layers /// - `pooler`: linear layer applied to the first element of the sequence (*[MASK]* token) /// - `is_decoder`: Flag indicating if the model is used as a decoder. If set to true, a causal mask will be applied to hide future positions that should not be attended to. pub struct BertModel<T: BertEmbedding> { embeddings: T, encoder: BertEncoder, pooler: BertPooler, is_decoder: bool, } /// Defines the implementation of the BertModel. The BERT model shares many similarities with RoBERTa, main difference being the embeddings. /// Therefore the forward pass of the model is shared and the type of embedding used is abstracted away. This allows to create /// `BertModel<RobertaEmbeddings>` or `BertModel<BertEmbeddings>` for each model type. impl<T: BertEmbedding> BertModel<T> { /// Build a new `BertModel` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BERT model /// * `config` - `BertConfig` object defining the model architecture and decoder status /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertEmbeddings, BertModel}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert: BertModel<BertEmbeddings> = BertModel::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertModel<T> where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let is_decoder = match config.is_decoder { Some(value) => value, None => false, }; let embeddings = T::new(p / "embeddings", config); let encoder = BertEncoder::new(p / "encoder", config); let pooler = BertPooler::new(p / "pooler", config); BertModel { embeddings, encoder, pooler, is_decoder, } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Optional input tensor of shape (*batch size*, *sequence_length*). If None, pre-computed embeddings must be provided (see `input_embeds`) /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` - Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `input_embeds` - Optional pre-computed input embeddings of shape (*batch size*, *sequence_length*, *hidden_size*). If None, input ids must be provided (see `input_ids`) /// * `encoder_hidden_states` - Optional encoder hidden state of shape (*batch size*, *encoder_sequence_length*, *hidden_size*). If the model is defined as a decoder and the `encoder_hidden_states` is not None, used in the cross-attention layer as keys and values (query from the decoder). /// * `encoder_mask` - Optional encoder attention mask of shape (*batch size*, *encoder_sequence_length*). If the model is defined as a decoder and the `encoder_hidden_states` is not None, used to mask encoder values. Positions with value 0 will be masked. /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `output` - `Tensor` of shape (*batch size*, *sequence_length*, *hidden_size*) /// * `pooled_output` - `Tensor` of shape (*batch size*, *hidden_size*) /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertModel, BertConfig, BertEmbeddings}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model: BertModel<BertEmbeddings> = BertModel::new(&vs.root(), &config); /// let (batch_size, sequence_length) = (64, 128); /// let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[batch_size, sequence_length], true); /// /// let (output, pooled_output, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model /// .forward_t( /// Some(input_tensor), /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// None, /// &None, /// &None, /// false, /// ) /// .unwrap() /// }); /// ``` pub fn forward_t( &self, input_ids: Option<Tensor>, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, input_embeds: Option<Tensor>, encoder_hidden_states: &Option<Tensor>, encoder_mask: &Option<Tensor>, train: bool, ) -> Result<(Tensor, Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>), &'static str> { let (input_shape, device) = match &input_ids { Some(input_value) => match &input_embeds { Some(_) => { return Err("Only one of input ids or input embeddings may be set"); } None => (input_value.size(), input_value.device()), }, None => match &input_embeds { Some(embeds) => (vec![embeds.size()[0], embeds.size()[1]], embeds.device()), None => { return Err("At least one of input ids or input embeddings must be set"); } }, }; let mask = match mask { Some(value) => value, None => Tensor::ones(&input_shape, (Kind::Int64, device)), }; let extended_attention_mask = match mask.dim() { 3 => mask.unsqueeze(1), 2 => { if self.is_decoder { let seq_ids = Tensor::arange(input_shape[1], (Float, device)); let causal_mask = seq_ids.unsqueeze(0).unsqueeze(0).repeat(&vec![ input_shape[0], input_shape[1], 1, ]); let causal_mask = causal_mask.le1(&seq_ids.unsqueeze(0).unsqueeze(-1)); causal_mask * mask.unsqueeze(1).unsqueeze(1) } else { mask.unsqueeze(1).unsqueeze(1) } } _ => { return Err("Invalid attention mask dimension, must be 2 or 3"); } }; let extended_attention_mask: Tensor = (extended_attention_mask.ones_like() - extended_attention_mask) * -10000.0; let encoder_extended_attention_mask: Option<Tensor> = if self.is_decoder & encoder_hidden_states.is_some() { let encoder_hidden_states = encoder_hidden_states.as_ref().unwrap(); let encoder_hidden_states_shape = encoder_hidden_states.size(); let encoder_mask = match encoder_mask { Some(value) => value.copy(), None => Tensor::ones( &[ encoder_hidden_states_shape[0], encoder_hidden_states_shape[1], ], (Kind::Int64, device), ), }; match encoder_mask.dim() { 2 => Some(encoder_mask.unsqueeze(1).unsqueeze(1)), 3 => Some(encoder_mask.unsqueeze(1)), _ => { return Err("Invalid encoder attention mask dimension, must be 2 or 3"); } } } else { None }; let embedding_output = match self.embeddings.forward_t( input_ids, token_type_ids, position_ids, input_embeds, train, ) { Ok(value) => value, Err(e) => { return Err(e); } }; let (hidden_state, all_hidden_states, all_attentions) = self.encoder.forward_t( &embedding_output, &Some(extended_attention_mask), encoder_hidden_states, &encoder_extended_attention_mask, train, ); let pooled_output = self.pooler.forward(&hidden_state); Ok(( hidden_state, pooled_output, all_hidden_states, all_attentions, )) } } pub struct BertPredictionHeadTransform { dense: nn::Linear, activation: Box<dyn Fn(&Tensor) -> Tensor>, layer_norm: nn::LayerNorm, } impl BertPredictionHeadTransform { pub fn new<'p, P>(p: P, config: &BertConfig) -> BertPredictionHeadTransform where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let dense = nn::linear( p / "dense", config.hidden_size, config.hidden_size, Default::default(), ); let activation = Box::new(match &config.hidden_act { Activation::gelu => _gelu, Activation::relu => _relu, Activation::mish => _mish, }); let layer_norm_config = nn::LayerNormConfig { eps: 1e-12, ..Default::default() }; let layer_norm = nn::layer_norm(p / "LayerNorm", vec![config.hidden_size], layer_norm_config); BertPredictionHeadTransform { dense, activation, layer_norm, } } pub fn forward(&self, hidden_states: &Tensor) -> Tensor { ((&self.activation)(&hidden_states.apply(&self.dense))).apply(&self.layer_norm) } } pub struct BertLMPredictionHead { transform: BertPredictionHeadTransform, decoder: LinearNoBias, bias: Tensor, } impl BertLMPredictionHead { pub fn new<'p, P>(p: P, config: &BertConfig) -> BertLMPredictionHead where P: Borrow<nn::Path<'p>>, { let p = p.borrow() / "predictions"; let transform = BertPredictionHeadTransform::new(&p / "transform", config); let decoder = linear_no_bias( &p / "decoder", config.hidden_size, config.vocab_size, Default::default(), ); let bias = p.var("bias", &[config.vocab_size], Init::KaimingUniform); BertLMPredictionHead { transform, decoder, bias, } } pub fn forward(&self, hidden_states: &Tensor) -> Tensor { self.transform.forward(&hidden_states).apply(&self.decoder) + &self.bias } } /// # BERT for masked language model /// Base BERT model with a masked language model head to predict missing tokens, for example `"Looks like one [MASK] is missing" -> "person"` /// It is made of the following blocks: /// - `bert`: Base BertModel /// - `cls`: BERT LM prediction head pub struct BertForMaskedLM { bert: BertModel<BertEmbeddings>, cls: BertLMPredictionHead, } impl BertForMaskedLM { /// Build a new `BertForMaskedLM` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BertForMaskedLM model /// * `config` - `BertConfig` object defining the model architecture and vocab size /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertForMaskedLM}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert = BertForMaskedLM::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertForMaskedLM where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let bert = BertModel::new(p / "bert", config); let cls = BertLMPredictionHead::new(p / "cls", config); BertForMaskedLM { bert, cls } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Optional input tensor of shape (*batch size*, *sequence_length*). If None, pre-computed embeddings must be provided (see *input_embeds*) /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` -Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `input_embeds` - Optional pre-computed input embeddings of shape (*batch size*, *sequence_length*, *hidden_size*). If None, input ids must be provided (see *input_ids*) /// * `encoder_hidden_states` - Optional encoder hidden state of shape (*batch size*, *encoder_sequence_length*, *hidden_size*). If the model is defined as a decoder and the *encoder_hidden_states* is not None, used in the cross-attention layer as keys and values (query from the decoder). /// * `encoder_mask` - Optional encoder attention mask of shape (*batch size*, *encoder_sequence_length*). If the model is defined as a decoder and the *encoder_hidden_states* is not None, used to mask encoder values. Positions with value 0 will be masked. /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `output` - `Tensor` of shape (*batch size*, *num_labels*, *vocab_size*) /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertForMaskedLM, BertConfig}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model = BertForMaskedLM::new(&vs.root(), &config); /// let (batch_size, sequence_length) = (64, 128); /// let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[batch_size, sequence_length], true); /// /// let (output, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model.forward_t( /// Some(input_tensor), /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// None, /// &None, /// &None, /// false, /// ) /// }); /// ``` pub fn forward_t( &self, input_ids: Option<Tensor>, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, input_embeds: Option<Tensor>, encoder_hidden_states: &Option<Tensor>, encoder_mask: &Option<Tensor>, train: bool, ) -> (Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>) { let (hidden_state, _, all_hidden_states, all_attentions) = self .bert .forward_t( input_ids, mask, token_type_ids, position_ids, input_embeds, encoder_hidden_states, encoder_mask, train, ) .unwrap(); let prediction_scores = self.cls.forward(&hidden_state); (prediction_scores, all_hidden_states, all_attentions) } } /// # BERT for sequence classification /// Base BERT model with a classifier head to perform sentence or document-level classification /// It is made of the following blocks: /// - `bert`: Base BertModel /// - `classifier`: BERT linear layer for classification pub struct BertForSequenceClassification { bert: BertModel<BertEmbeddings>, dropout: Dropout, classifier: nn::Linear, } impl BertForSequenceClassification { /// Build a new `BertForSequenceClassification` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BertForSequenceClassification model /// * `config` - `BertConfig` object defining the model architecture and number of classes /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertForSequenceClassification}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert = BertForSequenceClassification::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertForSequenceClassification where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let bert = BertModel::new(p / "bert", config); let dropout = Dropout::new(config.hidden_dropout_prob); let num_labels = config .id2label .as_ref() .expect("num_labels not provided in configuration") .len() as i64; let classifier = nn::linear( p / "classifier", config.hidden_size, num_labels, Default::default(), ); BertForSequenceClassification { bert, dropout, classifier, } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Optional input tensor of shape (*batch size*, *sequence_length*). If None, pre-computed embeddings must be provided (see `input_embeds`) /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` -Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `input_embeds` - Optional pre-computed input embeddings of shape (*batch size*, *sequence_length*, *hidden_size*). If None, input ids must be provided (see `input_ids`) /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `labels` - `Tensor` of shape (*batch size*, *num_labels*) /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertForSequenceClassification, BertConfig}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model = BertForSequenceClassification::new(&vs.root(), &config); /// let (batch_size, sequence_length) = (64, 128); /// let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[batch_size, sequence_length], true); /// /// let (labels, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model.forward_t( /// Some(input_tensor), /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// None, /// false, /// ) /// }); /// ``` pub fn forward_t( &self, input_ids: Option<Tensor>, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, input_embeds: Option<Tensor>, train: bool, ) -> (Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>) { let (_, pooled_output, all_hidden_states, all_attentions) = self .bert .forward_t( input_ids, mask, token_type_ids, position_ids, input_embeds, &None, &None, train, ) .unwrap(); let output = pooled_output .apply_t(&self.dropout, train) .apply(&self.classifier); (output, all_hidden_states, all_attentions) } } /// # BERT for multiple choices /// Multiple choices model using a BERT base model and a linear classifier. /// Input should be in the form `[CLS] Context [SEP] Possible choice [SEP]`. The choice is made along the batch axis, /// assuming all elements of the batch are alternatives to be chosen from for a given context. /// It is made of the following blocks: /// - `bert`: Base BertModel /// - `classifier`: Linear layer for multiple choices pub struct BertForMultipleChoice { bert: BertModel<BertEmbeddings>, dropout: Dropout, classifier: nn::Linear, } impl BertForMultipleChoice { /// Build a new `BertForMultipleChoice` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BertForMultipleChoice model /// * `config` - `BertConfig` object defining the model architecture /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertForMultipleChoice}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert = BertForMultipleChoice::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertForMultipleChoice where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let bert = BertModel::new(p / "bert", config); let dropout = Dropout::new(config.hidden_dropout_prob); let classifier = nn::linear(p / "classifier", config.hidden_size, 1, Default::default()); BertForMultipleChoice { bert, dropout, classifier, } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Input tensor of shape (*batch size*, *sequence_length*). /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` -Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `output` - `Tensor` of shape (*1*, *batch size*) containing the logits for each of the alternatives given /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertForMultipleChoice, BertConfig}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model = BertForMultipleChoice::new(&vs.root(), &config); /// let (num_choices, sequence_length) = (3, 128); /// let input_tensor = Tensor::rand(&[num_choices, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[num_choices, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[num_choices, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[num_choices, sequence_length], true); /// /// let (choices, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model.forward_t( /// input_tensor, /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// false, /// ) /// }); /// ``` pub fn forward_t( &self, input_ids: Tensor, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, train: bool, ) -> (Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>) { let num_choices = input_ids.size()[1]; let input_ids = input_ids.view((-1, *input_ids.size().last().unwrap())); let mask = match mask { Some(value) => Some(value.view((-1, *value.size().last().unwrap()))), None => None, }; let token_type_ids = match token_type_ids { Some(value) => Some(value.view((-1, *value.size().last().unwrap()))), None => None, }; let position_ids = match position_ids { Some(value) => Some(value.view((-1, *value.size().last().unwrap()))), None => None, }; let (_, pooled_output, all_hidden_states, all_attentions) = self .bert .forward_t( Some(input_ids), mask, token_type_ids, position_ids, None, &None, &None, train, ) .unwrap(); let output = pooled_output .apply_t(&self.dropout, train) .apply(&self.classifier) .view((-1, num_choices)); (output, all_hidden_states, all_attentions) } } /// # BERT for token classification (e.g. NER, POS) /// Token-level classifier predicting a label for each token provided. Note that because of wordpiece tokenization, the labels predicted are /// not necessarily aligned with words in the sentence. /// It is made of the following blocks: /// - `bert`: Base BertModel /// - `classifier`: Linear layer for token classification pub struct BertForTokenClassification { bert: BertModel<BertEmbeddings>, dropout: Dropout, classifier: nn::Linear, } impl BertForTokenClassification { /// Build a new `BertForTokenClassification` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BertForTokenClassification model /// * `config` - `BertConfig` object defining the model architecture, number of output labels and label mapping /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertForTokenClassification}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert = BertForTokenClassification::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertForTokenClassification where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let bert = BertModel::new(p / "bert", config); let dropout = Dropout::new(config.hidden_dropout_prob); let num_labels = config .id2label .as_ref() .expect("num_labels not provided in configuration") .len() as i64; let classifier = nn::linear( p / "classifier", config.hidden_size, num_labels, Default::default(), ); BertForTokenClassification { bert, dropout, classifier, } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Optional input tensor of shape (*batch size*, *sequence_length*). If None, pre-computed embeddings must be provided (see `input_embeds`) /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` -Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `input_embeds` - Optional pre-computed input embeddings of shape (*batch size*, *sequence_length*, *hidden_size*). If None, input ids must be provided (see `input_ids`) /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `output` - `Tensor` of shape (*batch size*, *sequence_length*, *num_labels*) containing the logits for each of the input tokens and classes /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertForTokenClassification, BertConfig}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model = BertForTokenClassification::new(&vs.root(), &config); /// let (batch_size, sequence_length) = (64, 128); /// let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[batch_size, sequence_length], true); /// /// let (token_labels, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model.forward_t( /// Some(input_tensor), /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// None, /// false, /// ) /// }); /// ``` pub fn forward_t( &self, input_ids: Option<Tensor>, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, input_embeds: Option<Tensor>, train: bool, ) -> (Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>) { let (hidden_state, _, all_hidden_states, all_attentions) = self .bert .forward_t( input_ids, mask, token_type_ids, position_ids, input_embeds, &None, &None, train, ) .unwrap(); let sequence_output = hidden_state .apply_t(&self.dropout, train) .apply(&self.classifier); (sequence_output, all_hidden_states, all_attentions) } } /// # BERT for question answering /// Extractive question-answering model based on a BERT language model. Identifies the segment of a context that answers a provided question. /// Please note that a significant amount of pre- and post-processing is required to perform end-to-end question answering. /// See the question answering pipeline (also provided in this crate) for more details. /// It is made of the following blocks: /// - `bert`: Base BertModel /// - `qa_outputs`: Linear layer for question answering pub struct BertForQuestionAnswering { bert: BertModel<BertEmbeddings>, qa_outputs: nn::Linear, } impl BertForQuestionAnswering { /// Build a new `BertForQuestionAnswering` /// /// # Arguments /// /// * `p` - Variable store path for the root of the BertForQuestionAnswering model /// * `config` - `BertConfig` object defining the model architecture /// /// # Example /// /// ```no_run /// use rust_bert::bert::{BertConfig, BertForQuestionAnswering}; /// use rust_bert::Config; /// use std::path::Path; /// use tch::{nn, Device}; /// /// let config_path = Path::new("path/to/config.json"); /// let device = Device::Cpu; /// let p = nn::VarStore::new(device); /// let config = BertConfig::from_file(config_path); /// let bert = BertForQuestionAnswering::new(&p.root() / "bert", &config); /// ``` pub fn new<'p, P>(p: P, config: &BertConfig) -> BertForQuestionAnswering where P: Borrow<nn::Path<'p>>, { let p = p.borrow(); let bert = BertModel::new(p / "bert", config); let num_labels = 2; let qa_outputs = nn::linear( p / "qa_outputs", config.hidden_size, num_labels, Default::default(), ); BertForQuestionAnswering { bert, qa_outputs } } /// Forward pass through the model /// /// # Arguments /// /// * `input_ids` - Optional input tensor of shape (*batch size*, *sequence_length*). If None, pre-computed embeddings must be provided (see `input_embeds`) /// * `mask` - Optional mask of shape (*batch size*, *sequence_length*). Masked position have value 0, non-masked value 1. If None set to 1 /// * `token_type_ids` -Optional segment id of shape (*batch size*, *sequence_length*). Convention is value of 0 for the first sentence (incl. *[SEP]*) and 1 for the second sentence. If None set to 0. /// * `position_ids` - Optional position ids of shape (*batch size*, *sequence_length*). If None, will be incremented from 0. /// * `input_embeds` - Optional pre-computed input embeddings of shape (*batch size*, *sequence_length*, *hidden_size*). If None, input ids must be provided (see `input_ids`) /// * `train` - boolean flag to turn on/off the dropout layers in the model. Should be set to false for inference. /// /// # Returns /// /// * `start_scores` - `Tensor` of shape (*batch size*, *sequence_length*) containing the logits for start of the answer /// * `end_scores` - `Tensor` of shape (*batch size*, *sequence_length*) containing the logits for end of the answer /// * `hidden_states` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// * `attentions` - `Option<Vec<Tensor>>` of length *num_hidden_layers* with shape (*batch size*, *sequence_length*, *hidden_size*) /// /// # Example /// /// ```no_run /// # use rust_bert::bert::{BertForQuestionAnswering, BertConfig}; /// # use tch::{nn, Device, Tensor, no_grad}; /// # use rust_bert::Config; /// # use std::path::Path; /// # use tch::kind::Kind::Int64; /// # let config_path = Path::new("path/to/config.json"); /// # let device = Device::Cpu; /// # let vs = nn::VarStore::new(device); /// # let config = BertConfig::from_file(config_path); /// # let bert_model = BertForQuestionAnswering::new(&vs.root(), &config); /// let (batch_size, sequence_length) = (64, 128); /// let input_tensor = Tensor::rand(&[batch_size, sequence_length], (Int64, device)); /// let mask = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let token_type_ids = Tensor::zeros(&[batch_size, sequence_length], (Int64, device)); /// let position_ids = Tensor::arange(sequence_length, (Int64, device)) /// .expand(&[batch_size, sequence_length], true); /// /// let (start_scores, end_scores, all_hidden_states, all_attentions) = no_grad(|| { /// bert_model.forward_t( /// Some(input_tensor), /// Some(mask), /// Some(token_type_ids), /// Some(position_ids), /// None, /// false, /// ) /// }); /// ``` pub fn forward_t( &self, input_ids: Option<Tensor>, mask: Option<Tensor>, token_type_ids: Option<Tensor>, position_ids: Option<Tensor>, input_embeds: Option<Tensor>, train: bool, ) -> (Tensor, Tensor, Option<Vec<Tensor>>, Option<Vec<Tensor>>) { let (hidden_state, _, all_hidden_states, all_attentions) = self .bert .forward_t( input_ids, mask, token_type_ids, position_ids, input_embeds, &None, &None, train, ) .unwrap(); let sequence_output = hidden_state.apply(&self.qa_outputs); let logits = sequence_output.split(1, -1); let (start_logits, end_logits) = (&logits[0], &logits[1]); let start_logits = start_logits.squeeze1(-1); let end_logits = end_logits.squeeze1(-1); (start_logits, end_logits, all_hidden_states, all_attentions) } }