diff --git a/README.md b/README.md index aa42640226..84569c6c6b 100644 --- a/README.md +++ b/README.md @@ -301,6 +301,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel. +1. **[MVP](https://huggingface.co/docs/transformers/main/model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen. 1. **[Nezha](https://huggingface.co/docs/transformers/main/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu. 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al. diff --git a/README_ko.md b/README_ko.md index b65577c3b3..f297831f40 100644 --- a/README_ko.md +++ b/README_ko.md @@ -282,6 +282,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel. +1. **[MVP](https://huggingface.co/docs/transformers/main/model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen. 1. **[Nezha](https://huggingface.co/docs/transformers/main/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu. 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al. @@ -370,4 +371,4 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6", pages = "38--45" } -``` \ No newline at end of file +``` diff --git a/README_zh-hans.md b/README_zh-hans.md index 10816882a5..6923a6c5de 100644 --- a/README_zh-hans.md +++ b/README_zh-hans.md @@ -306,6 +306,7 @@ conda install -c huggingface transformers 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (来自 CMU/Google Brain) 伴随论文 [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) 由 Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou 发布。 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (来自 Microsoft Research) 伴随论文 [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) 由 Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu 发布。 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (来自 Google AI) 伴随论文 [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) 由 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel 发布。 +1. **[MVP](https://huggingface.co/docs/transformers/main/model_doc/mvp)** (来自 中国人民大学 AI Box) 伴随论文 [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) 由 Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen 发布。 1. **[Nezha](https://huggingface.co/docs/transformers/main/model_doc/nezha)** (来自华为诺亚方舟实验室) 伴随论文 [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) 由 Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu 发布。 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (来自 the University of Wisconsin - Madison) 伴随论文 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 由 Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 发布。 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (来自 Meta AI) 伴随论文 [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) 由 Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al 发布。 diff --git a/README_zh-hant.md b/README_zh-hant.md index 8607a40b3a..54d9a11129 100644 --- a/README_zh-hant.md +++ b/README_zh-hant.md @@ -318,6 +318,7 @@ conda install -c huggingface transformers 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel. +1. **[MVP](https://huggingface.co/docs/transformers/main/model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen. 1. **[Nezha](https://huggingface.co/docs/transformers/main/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu. 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al. @@ -407,4 +408,4 @@ conda install -c huggingface transformers url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6", pages = "38--45" } -``` \ No newline at end of file +``` diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 280285371b..fbcf68e39b 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -308,6 +308,8 @@ title: MPNet - local: model_doc/mt5 title: MT5 + - local: model_doc/mvp + title: MVP - local: model_doc/nezha title: NEZHA - local: model_doc/nystromformer diff --git a/docs/source/en/index.mdx b/docs/source/en/index.mdx index 89072ae5de..1346414bed 100644 --- a/docs/source/en/index.mdx +++ b/docs/source/en/index.mdx @@ -124,6 +124,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret 1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. 1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel. +1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen. 1. **[Nezha](model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu. 1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh. 1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al. @@ -253,6 +254,7 @@ Flax), PyTorch, and/or TensorFlow. | MobileBERT | ✅ | ✅ | ✅ | ✅ | ❌ | | MPNet | ✅ | ✅ | ✅ | ✅ | ❌ | | MT5 | ✅ | ✅ | ✅ | ✅ | ✅ | +| MVP | ✅ | ✅ | ✅ | ❌ | ❌ | | Nezha | ❌ | ❌ | ✅ | ❌ | ❌ | | Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ | | OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ | diff --git a/docs/source/en/model_doc/mvp.mdx b/docs/source/en/model_doc/mvp.mdx new file mode 100644 index 0000000000..6fae8c7311 --- /dev/null +++ b/docs/source/en/model_doc/mvp.mdx @@ -0,0 +1,138 @@ + + +# MVP + +## Overview + +The MVP model was proposed in [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen. + + +According to the abstract, + +- MVP follows a standard Transformer encoder-decoder architecture. +- MVP is supervised pre-trained using labeled datasets. +- MVP also has task-specific soft prompts to stimulate the model's capacity in performing a certain task. +- MVP is specially designed for natural language generation and can be adapted to a wide range of generation tasks, including but not limited to summarization, data-to-text generation, open-ended dialogue system, story generation, question answering, question generation, task-oriented dialogue system, commonsense generation, paraphrase generation, text style transfer, and text simplification. Our model can also be adapted to natural language understanding tasks such as sequence classification and (extractive) question answering. + +Tips: +- We have released a series of models [here](https://huggingface.co/models?filter=mvp), including MVP, MVP with task-specific prompts, and multi-task pre-trained variants. +- If you want to use a model without prompts (standard Transformer), you can load it through `MvpForConditionalGeneration.from_pretrained('RUCAIBox/mvp')`. +- If you want to use a model with task-specific prompts, such as summarization, you can load it through `MvpForConditionalGeneration.from_pretrained('RUCAIBox/mvp-summarization')`. +- Our model supports lightweight prompt tuning following [Prefix-tuning](https://arxiv.org/abs/2101.00190) with method `set_lightweight_tuning()`. + +This model was contributed by [Tianyi Tang](https://huggingface.co/StevenTang). The detailed information and instructions can be found [here](https://github.com/RUCAIBox/MVP). + +## Examples +For summarization, it is an example to use MVP and MVP with summarization-specific prompts. + +```python +>>> from transformers import MvpTokenizer, MvpForConditionalGeneration + +>>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") +>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp") +>>> model_with_prompt = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp-summarization") + +>>> inputs = tokenizer( +... "Summarize: You may want to stick it to your boss and leave your job, but don't do it if these are your reasons.", +... return_tensors="pt", +... ) +>>> generated_ids = model.generate(**inputs) +>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True) +["Why You Shouldn't Quit Your Job"] + +>>> generated_ids = model_with_prompt.generate(**inputs) +>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True) +["Don't do it if these are your reasons"] +``` + +For data-to-text generation, it is an example to use MVP and multi-task pre-trained variants. +```python +>>> from transformers import MvpTokenizerFast, MvpForConditionalGeneration + +>>> tokenizer = MvpTokenizerFast.from_pretrained("RUCAIBox/mvp") +>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp") +>>> model_with_mtl = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mtl-data-to-text") + +>>> inputs = tokenizer( +... "Describe the following data: Iron Man | instance of | Superhero [SEP] Stan Lee | creator | Iron Man", +... return_tensors="pt", +... ) +>>> generated_ids = model.generate(**inputs) +>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True) +['Stan Lee created the character of Iron Man, a fictional superhero appearing in American comic'] + +>>> generated_ids = model_with_mtl.generate(**inputs) +>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True) +['Iron Man is a fictional superhero appearing in American comic books published by Marvel Comics.'] +``` + +For lightweight tuning, *i.e.*, fixing the model and only tuning prompts, you can load MVP with randomly initialized prompts or with task-specific prompts. Our code also supports Prefix-tuning with BART following the [original paper](https://arxiv.org/abs/2101.00190). + +```python +>>> from transformers import MvpForConditionalGeneration + +>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp", use_prompt=True) +>>> # the number of trainable parameters (full tuning) +>>> sum(p.numel() for p in model.parameters() if p.requires_grad) +468116832 + +>>> # lightweight tuning with randomly initialized prompts +>>> model.set_lightweight_tuning() +>>> # the number of trainable parameters (lightweight tuning) +>>> sum(p.numel() for p in model.parameters() if p.requires_grad) +61823328 + +>>> # lightweight tuning with task-specific prompts +>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mtl-data-to-text") +>>> model.set_lightweight_tuning() +>>> # original lightweight Prefix-tuning +>>> model = MvpForConditionalGeneration.from_pretrained("facebook/bart-large", use_prompt=True) +>>> model.set_lightweight_tuning() +``` + +## MvpConfig + +[[autodoc]] MvpConfig + +## MvpTokenizer + +[[autodoc]] MvpTokenizer + +## MvpTokenizerFast + +[[autodoc]] MvpTokenizerFast + +## MvpModel + +[[autodoc]] MvpModel + - forward + +## MvpForConditionalGeneration + +[[autodoc]] MvpForConditionalGeneration + - forward + +## MvpForSequenceClassification + +[[autodoc]] MvpForSequenceClassification + - forward + +## MvpForQuestionAnswering + +[[autodoc]] MvpForQuestionAnswering + - forward + +## MvpForCausalLM + +[[autodoc]] MvpForCausalLM + - forward diff --git a/docs/source/en/serialization.mdx b/docs/source/en/serialization.mdx index f5c113e310..846bedff9d 100644 --- a/docs/source/en/serialization.mdx +++ b/docs/source/en/serialization.mdx @@ -677,4 +677,4 @@ torch.neuron.trace(model, [token_tensor, segments_tensors]) This change enables Neuron SDK to trace the model and optimize it to run in Inf1 instances. To learn more about AWS Neuron SDK features, tools, example tutorials and latest updates, -please see the [AWS NeuronSDK documentation](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html). \ No newline at end of file +please see the [AWS NeuronSDK documentation](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html). diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index 17670395f3..dabc7e2ea8 100755 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -263,6 +263,7 @@ _import_structure = { "models.mobilebert": ["MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MobileBertConfig", "MobileBertTokenizer"], "models.mpnet": ["MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "MPNetConfig", "MPNetTokenizer"], "models.mt5": ["MT5Config"], + "models.mvp": ["MvpConfig", "MvpTokenizer"], "models.nezha": ["NEZHA_PRETRAINED_CONFIG_ARCHIVE_MAP", "NezhaConfig"], "models.nystromformer": [ "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", @@ -540,6 +541,7 @@ else: _import_structure["models.mobilebert"].append("MobileBertTokenizerFast") _import_structure["models.mpnet"].append("MPNetTokenizerFast") _import_structure["models.mt5"].append("MT5TokenizerFast") + _import_structure["models.mvp"].append("MvpTokenizerFast") _import_structure["models.openai"].append("OpenAIGPTTokenizerFast") _import_structure["models.pegasus"].append("PegasusTokenizerFast") _import_structure["models.realm"].append("RealmTokenizerFast") @@ -826,6 +828,17 @@ else: "PretrainedBartModel", ] ) + _import_structure["models.mvp"].extend( + [ + "MVP_PRETRAINED_MODEL_ARCHIVE_LIST", + "MvpForCausalLM", + "MvpForConditionalGeneration", + "MvpForQuestionAnswering", + "MvpForSequenceClassification", + "MvpModel", + "MvpPreTrainedModel", + ] + ) _import_structure["models.beit"].extend( [ "BEIT_PRETRAINED_MODEL_ARCHIVE_LIST", @@ -2948,6 +2961,7 @@ if TYPE_CHECKING: from .models.mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig, MobileBertTokenizer from .models.mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig, MPNetTokenizer from .models.mt5 import MT5Config + from .models.mvp import MvpConfig, MvpTokenizer from .models.nezha import NEZHA_PRETRAINED_CONFIG_ARCHIVE_MAP, NezhaConfig from .models.nystromformer import NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, NystromformerConfig from .models.openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig, OpenAIGPTTokenizer @@ -3195,6 +3209,7 @@ if TYPE_CHECKING: from .models.mobilebert import MobileBertTokenizerFast from .models.mpnet import MPNetTokenizerFast from .models.mt5 import MT5TokenizerFast + from .models.mvp import MvpTokenizerFast from .models.openai import OpenAIGPTTokenizerFast from .models.pegasus import PegasusTokenizerFast from .models.realm import RealmTokenizerFast @@ -3925,6 +3940,15 @@ if TYPE_CHECKING: MPNetPreTrainedModel, ) from .models.mt5 import MT5EncoderModel, MT5ForConditionalGeneration, MT5Model + from .models.mvp import ( + MVP_PRETRAINED_MODEL_ARCHIVE_LIST, + MvpForCausalLM, + MvpForConditionalGeneration, + MvpForQuestionAnswering, + MvpForSequenceClassification, + MvpModel, + MvpPreTrainedModel, + ) from .models.nezha import ( NEZHA_PRETRAINED_MODEL_ARCHIVE_LIST, NezhaForMaskedLM, diff --git a/src/transformers/convert_slow_tokenizer.py b/src/transformers/convert_slow_tokenizer.py index ea59284782..9059854799 100644 --- a/src/transformers/convert_slow_tokenizer.py +++ b/src/transformers/convert_slow_tokenizer.py @@ -1033,6 +1033,7 @@ SLOW_TO_FAST_CONVERTERS = { "MBart50Tokenizer": MBart50Converter, "MPNetTokenizer": MPNetConverter, "MobileBertTokenizer": BertConverter, + "MvpTokenizer": RobertaConverter, "OpenAIGPTTokenizer": OpenAIGPTConverter, "PegasusTokenizer": PegasusConverter, "RealmTokenizer": BertConverter, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index 21ab46f5e6..07140cbf74 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -94,6 +94,7 @@ from . import ( mobilebert, mpnet, mt5, + mvp, nezha, nystromformer, openai, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index a8a12f6e30..b5ee08df8c 100644 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -92,6 +92,7 @@ CONFIG_MAPPING_NAMES = OrderedDict( ("mobilebert", "MobileBertConfig"), ("mpnet", "MPNetConfig"), ("mt5", "MT5Config"), + ("mvp", "MvpConfig"), ("nezha", "NezhaConfig"), ("nystromformer", "NystromformerConfig"), ("openai-gpt", "OpenAIGPTConfig"), @@ -208,6 +209,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict( ("mctct", "MCTCT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("megatron-bert", "MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("mpnet", "MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP"), + ("mvp", "MVP_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("nezha", "NEZHA_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("nystromformer", "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("openai-gpt", "OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP"), @@ -335,6 +337,7 @@ MODEL_NAMES_MAPPING = OrderedDict( ("mobilebert", "MobileBERT"), ("mpnet", "MPNet"), ("mt5", "MT5"), + ("mvp", "MVP"), ("nezha", "Nezha"), ("nystromformer", "Nyströmformer"), ("openai-gpt", "OpenAI GPT"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index 90171f05b9..9bd12f8a87 100644 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -91,6 +91,7 @@ MODEL_MAPPING_NAMES = OrderedDict( ("mobilebert", "MobileBertModel"), ("mpnet", "MPNetModel"), ("mt5", "MT5Model"), + ("mvp", "MvpModel"), ("nezha", "NezhaModel"), ("nystromformer", "NystromformerModel"), ("openai-gpt", "OpenAIGPTModel"), @@ -169,6 +170,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict( ("megatron-bert", "MegatronBertForPreTraining"), ("mobilebert", "MobileBertForPreTraining"), ("mpnet", "MPNetForMaskedLM"), + ("mvp", "MvpForConditionalGeneration"), ("nezha", "NezhaForPreTraining"), ("openai-gpt", "OpenAIGPTLMHeadModel"), ("retribert", "RetriBertModel"), @@ -229,6 +231,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict( ("megatron-bert", "MegatronBertForCausalLM"), ("mobilebert", "MobileBertForMaskedLM"), ("mpnet", "MPNetForMaskedLM"), + ("mvp", "MvpForConditionalGeneration"), ("nezha", "NezhaForMaskedLM"), ("nystromformer", "NystromformerForMaskedLM"), ("openai-gpt", "OpenAIGPTLMHeadModel"), @@ -275,6 +278,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict( ("marian", "MarianForCausalLM"), ("mbart", "MBartForCausalLM"), ("megatron-bert", "MegatronBertForCausalLM"), + ("mvp", "MvpForCausalLM"), ("openai-gpt", "OpenAIGPTLMHeadModel"), ("opt", "OPTForCausalLM"), ("pegasus", "PegasusForCausalLM"), @@ -397,6 +401,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict( ("megatron-bert", "MegatronBertForMaskedLM"), ("mobilebert", "MobileBertForMaskedLM"), ("mpnet", "MPNetForMaskedLM"), + ("mvp", "MvpForConditionalGeneration"), ("nezha", "NezhaForMaskedLM"), ("nystromformer", "NystromformerForMaskedLM"), ("perceiver", "PerceiverForMaskedLM"), @@ -438,6 +443,7 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict( ("marian", "MarianMTModel"), ("mbart", "MBartForConditionalGeneration"), ("mt5", "MT5ForConditionalGeneration"), + ("mvp", "MvpForConditionalGeneration"), ("pegasus", "PegasusForConditionalGeneration"), ("plbart", "PLBartForConditionalGeneration"), ("prophetnet", "ProphetNetForConditionalGeneration"), @@ -487,6 +493,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("megatron-bert", "MegatronBertForSequenceClassification"), ("mobilebert", "MobileBertForSequenceClassification"), ("mpnet", "MPNetForSequenceClassification"), + ("mvp", "MvpForSequenceClassification"), ("nezha", "NezhaForSequenceClassification"), ("nystromformer", "NystromformerForSequenceClassification"), ("openai-gpt", "OpenAIGPTForSequenceClassification"), @@ -538,6 +545,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict( ("megatron-bert", "MegatronBertForQuestionAnswering"), ("mobilebert", "MobileBertForQuestionAnswering"), ("mpnet", "MPNetForQuestionAnswering"), + ("mvp", "MvpForQuestionAnswering"), ("nezha", "NezhaForQuestionAnswering"), ("nystromformer", "NystromformerForQuestionAnswering"), ("qdqbert", "QDQBertForQuestionAnswering"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index c5fb57626b..d2f3477616 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -175,6 +175,7 @@ else: "MT5TokenizerFast" if is_tokenizers_available() else None, ), ), + ("mvp", ("MvpTokenizer", "MvpTokenizerFast" if is_tokenizers_available() else None)), ("nezha", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)), ( "nystromformer", diff --git a/src/transformers/models/mvp/__init__.py b/src/transformers/models/mvp/__init__.py new file mode 100644 index 0000000000..865b958d39 --- /dev/null +++ b/src/transformers/models/mvp/__init__.py @@ -0,0 +1,83 @@ +# flake8: noqa +# There's no way to ignore "F401 '...' imported but unused" warnings in this +# module, but to preserve other warnings. So, don't check this module at all. + +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# 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. +from typing import TYPE_CHECKING + +from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available + + +_import_structure = { + "configuration_mvp": ["MVP_PRETRAINED_CONFIG_ARCHIVE_MAP", "MvpConfig", "MvpOnnxConfig"], + "tokenization_mvp": ["MvpTokenizer"], +} + +try: + if not is_tokenizers_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["tokenization_mvp_fast"] = ["MvpTokenizerFast"] + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["modeling_mvp"] = [ + "MVP_PRETRAINED_MODEL_ARCHIVE_LIST", + "MvpForCausalLM", + "MvpForConditionalGeneration", + "MvpForQuestionAnswering", + "MvpForSequenceClassification", + "MvpModel", + "MvpPreTrainedModel", + ] + +if TYPE_CHECKING: + from .configuration_mvp import MVP_PRETRAINED_CONFIG_ARCHIVE_MAP, MvpConfig, MvpOnnxConfig + from .tokenization_mvp import MvpTokenizer + + try: + if not is_tokenizers_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .tokenization_mvp_fast import MvpTokenizerFast + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .modeling_mvp import ( + MVP_PRETRAINED_MODEL_ARCHIVE_LIST, + MvpForCausalLM, + MvpForConditionalGeneration, + MvpForQuestionAnswering, + MvpForSequenceClassification, + MvpModel, + MvpPreTrainedModel, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) diff --git a/src/transformers/models/mvp/configuration_mvp.py b/src/transformers/models/mvp/configuration_mvp.py new file mode 100644 index 0000000000..63a006b8e4 --- /dev/null +++ b/src/transformers/models/mvp/configuration_mvp.py @@ -0,0 +1,182 @@ +# coding=utf-8 +# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" MVP model configuration""" +import warnings + +from ...configuration_utils import PretrainedConfig +from ...utils import logging + + +logger = logging.get_logger(__name__) + +MVP_PRETRAINED_CONFIG_ARCHIVE_MAP = { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/config.json", +} + + +class MvpConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`MvpModel`]. It is used to instantiate a MVP model + according to the specified arguments, defining the model architecture. Instantiating a configuration with the + defaults will yield a similar configuration to that of the MVP [RUCAIBox/mvp](https://huggingface.co/RUCAIBox/mvp) + architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + + Args: + vocab_size (`int`, *optional*, defaults to 50267): + Vocabulary size of the MVP model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`MvpModel`]. + d_model (`int`, *optional*, defaults to 1024): + Dimensionality of the layers and the pooler layer. + encoder_layers (`int`, *optional*, defaults to 12): + Number of encoder layers. + decoder_layers (`int`, *optional*, defaults to 12): + Number of decoder layers. + encoder_attention_heads (`int`, *optional*, defaults to 16): + Number of attention heads for each attention layer in the Transformer encoder. + decoder_attention_heads (`int`, *optional*, defaults to 16): + Number of attention heads for each attention layer in the Transformer decoder. + decoder_ffn_dim (`int`, *optional*, defaults to 4096): + Dimensionality of the "intermediate" (often named feed-forward) layer in decoder. + encoder_ffn_dim (`int`, *optional*, defaults to 4096): + Dimensionality of the "intermediate" (often named feed-forward) layer in decoder. + activation_function (`str` or `function`, *optional*, defaults to `"gelu"`): + The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, + `"relu"`, `"silu"` and `"gelu_new"` are supported. + dropout (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + activation_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for activations inside the fully connected layer. + classifier_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for classifier. + max_position_embeddings (`int`, *optional*, defaults to 1024): + The maximum sequence length that this model might ever be used with. Typically set this to something large + just in case (e.g., 512 or 1024 or 2048). + init_std (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + encoder_layerdrop (`float`, *optional*, defaults to 0.0): + The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) + for more details. + decoder_layerdrop (`float`, *optional*, defaults to 0.0): + The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) + for more details. + scale_embedding (`bool`, *optional*, defaults to `False`): + Scale embeddings by diving by sqrt(d_model). + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). + forced_eos_token_id (`int`, *optional*, defaults to 2): + The id of the token to force as the last generated token when `max_length` is reached. Usually set to + `eos_token_id`. + use_prompt (`bool`, *optional*, defaults to `False`): + Whether or not to use prompt. + prompt_length (`int`, *optional*, defaults to 100): + The length of prompt. + prompt_mid_dim (`int`, *optional*, defaults to 800): + Dimensionality of the "intermediate" layer in prompt. + Example: + + ```python + >>> from transformers import MvpModel, MvpConfig + + >>> # Initializing a MVP RUCAIBox/mvp style configuration + >>> configuration = MvpConfig() + + >>> # Initializing a model from the RUCAIBox/mvp style configuration + >>> model = MvpModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "mvp" + keys_to_ignore_at_inference = ["past_key_values"] + attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"} + + def __init__( + self, + vocab_size=50267, + max_position_embeddings=1024, + encoder_layers=12, + encoder_ffn_dim=4096, + encoder_attention_heads=16, + decoder_layers=12, + decoder_ffn_dim=4096, + decoder_attention_heads=16, + encoder_layerdrop=0.0, + decoder_layerdrop=0.0, + activation_function="gelu", + d_model=1024, + dropout=0.1, + attention_dropout=0.0, + activation_dropout=0.0, + init_std=0.02, + classifier_dropout=0.0, + scale_embedding=False, + use_cache=True, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + is_encoder_decoder=True, + decoder_start_token_id=2, + forced_eos_token_id=2, + use_prompt=False, + prompt_length=100, + prompt_mid_dim=800, + **kwargs + ): + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.d_model = d_model + self.encoder_ffn_dim = encoder_ffn_dim + self.encoder_layers = encoder_layers + self.encoder_attention_heads = encoder_attention_heads + self.decoder_ffn_dim = decoder_ffn_dim + self.decoder_layers = decoder_layers + self.decoder_attention_heads = decoder_attention_heads + self.dropout = dropout + self.attention_dropout = attention_dropout + self.activation_dropout = activation_dropout + self.activation_function = activation_function + self.init_std = init_std + self.encoder_layerdrop = encoder_layerdrop + self.decoder_layerdrop = decoder_layerdrop + self.classifier_dropout = classifier_dropout + self.use_cache = use_cache + self.num_hidden_layers = encoder_layers + self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True + self.use_prompt = use_prompt + self.prompt_length = prompt_length + self.prompt_mid_dim = prompt_mid_dim + + super().__init__( + pad_token_id=pad_token_id, + bos_token_id=bos_token_id, + eos_token_id=eos_token_id, + is_encoder_decoder=is_encoder_decoder, + decoder_start_token_id=decoder_start_token_id, + forced_eos_token_id=forced_eos_token_id, + **kwargs, + ) + + if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False): + self.forced_bos_token_id = self.bos_token_id + warnings.warn( + f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions. " + "The config can simply be saved and uploaded again to be fixed." + ) diff --git a/src/transformers/models/mvp/modeling_mvp.py b/src/transformers/models/mvp/modeling_mvp.py new file mode 100644 index 0000000000..d3d239c4cf --- /dev/null +++ b/src/transformers/models/mvp/modeling_mvp.py @@ -0,0 +1,2050 @@ +# coding=utf-8 +# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" PyTorch MVP model.""" +import copy +import math +import random +from typing import List, Optional, Tuple, Union + +import torch +import torch.utils.checkpoint +from torch import nn +from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss + +from ...activations import ACT2FN +from ...modeling_outputs import ( + BaseModelOutput, + BaseModelOutputWithPastAndCrossAttentions, + CausalLMOutputWithCrossAttentions, + Seq2SeqLMOutput, + Seq2SeqModelOutput, + Seq2SeqQuestionAnsweringModelOutput, + Seq2SeqSequenceClassifierOutput, +) +from ...modeling_utils import PreTrainedModel +from ...utils import ( + add_code_sample_docstrings, + add_end_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + logging, + replace_return_docstrings, +) +from .configuration_mvp import MvpConfig + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "RUCAIBox/mvp" +_CONFIG_FOR_DOC = "MvpConfig" +_TOKENIZER_FOR_DOC = "MvpTokenizer" + +# Base model docstring +_EXPECTED_OUTPUT_SHAPE = [1, 8, 1024] + +MVP_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "RUCAIBox/mvp", + "RUCAIBox/mvp-data-to-text", + "RUCAIBox/mvp-open-dialog", + "RUCAIBox/mvp-question-answering", + "RUCAIBox/mvp-question-generation", + "RUCAIBox/mvp-story", + "RUCAIBox/mvp-summarization", + "RUCAIBox/mvp-task-dialog", + "RUCAIBox/mtl-data-to-text", + "RUCAIBox/mtl-multi-task", + "RUCAIBox/mtl-open-dialog", + "RUCAIBox/mtl-question-answering", + "RUCAIBox/mtl-question-generation", + "RUCAIBox/mtl-story", + "RUCAIBox/mtl-summarization", + # See all MVP models at https://huggingface.co/models?filter=mvp +] + + +# Copied from transformers.models.bart.modeling_bart.shift_tokens_right +def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int): + """ + Shift input ids one token to the right. + """ + shifted_input_ids = input_ids.new_zeros(input_ids.shape) + shifted_input_ids[:, 1:] = input_ids[:, :-1].clone() + shifted_input_ids[:, 0] = decoder_start_token_id + + if pad_token_id is None: + raise ValueError("self.model.config.pad_token_id has to be defined.") + # replace possible -100 values in labels by `pad_token_id` + shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) + + return shifted_input_ids + + +# Copied from transformers.models.bart.modeling_bart._make_causal_mask +def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0): + """ + Make causal mask used for bi-directional self-attention. + """ + bsz, tgt_len = input_ids_shape + mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min)) + mask_cond = torch.arange(mask.size(-1)) + mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0) + mask = mask.to(dtype) + + if past_key_values_length > 0: + mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1) + return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length) + + +# Copied from transformers.models.bart.modeling_bart._expand_mask +def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None): + """ + Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. + """ + bsz, src_len = mask.size() + tgt_len = tgt_len if tgt_len is not None else src_len + + expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) + + inverted_mask = 1.0 - expanded_mask + + return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min) + + +# Copied from transformers.models.bart.modeling_bart.BartLearnedPositionalEmbedding with Bart->MVP +class MvpLearnedPositionalEmbedding(nn.Embedding): + """ + This module learns positional embeddings up to a fixed maximum size. + """ + + def __init__(self, num_embeddings: int, embedding_dim: int): + # MVP is set up so that if padding_idx is specified then offset the embedding ids by 2 + # and adjust num_embeddings appropriately. Other models don't have this hack + self.offset = 2 + super().__init__(num_embeddings + self.offset, embedding_dim) + + def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0): + """`input_ids_shape` is expected to be [bsz x seqlen].""" + bsz, seq_len = input_ids_shape[:2] + positions = torch.arange( + past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device + ) + return super().forward(positions + self.offset) + + +class MvpAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + is_decoder: bool = False, + bias: bool = True, + ): + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + + if (self.head_dim * num_heads) != self.embed_dim: + raise ValueError( + f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" + f" and `num_heads`: {num_heads})." + ) + self.scaling = self.head_dim**-0.5 + self.is_decoder = is_decoder + + self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + + def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): + return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() + + def forward( + self, + hidden_states: torch.Tensor, + key_value_states: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + attention_mask: Optional[torch.Tensor] = None, + layer_head_mask: Optional[torch.Tensor] = None, + attn_prompt: Optional[torch.Tensor] = None, + output_attentions: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + """Input shape: Batch x Time x Channel""" + + # if key_value_states are provided this layer is used as a cross-attention layer + # for the decoder + is_cross_attention = key_value_states is not None + + bsz, tgt_len, _ = hidden_states.size() + + # get query proj + query_states = self.q_proj(hidden_states) * self.scaling + # get key, value proj + if is_cross_attention and past_key_value is not None: + # reuse k,v, cross_attentions + key_states = past_key_value[0] + value_states = past_key_value[1] + elif is_cross_attention: + # cross_attentions + key_states = self._shape(self.k_proj(key_value_states), -1, bsz) + value_states = self._shape(self.v_proj(key_value_states), -1, bsz) + elif past_key_value is not None: + # reuse k, v, self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) + else: + # self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + + if self.is_decoder: + # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. + # Further calls to cross_attention layer can then reuse all cross-attention + # key/value_states (first "if" case) + # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of + # all previous decoder key/value_states. Further calls to uni-directional self-attention + # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) + # if encoder bi-directional self-attention `past_key_value` is always `None` + past_key_value = (key_states, value_states) + + if attn_prompt is not None: + key_states = torch.cat([attn_prompt[0].expand(bsz, -1, -1, -1), key_states], dim=2) + value_states = torch.cat([attn_prompt[1].expand(bsz, -1, -1, -1), value_states], dim=2) + if attention_mask is not None: + prompt_mask = torch.zeros(bsz, 1, tgt_len, attn_prompt[0].size(1)).to(attention_mask.device) + attention_mask = torch.cat([prompt_mask, attention_mask], dim=(-1)) + + proj_shape = (bsz * self.num_heads, -1, self.head_dim) + query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) + key_states = key_states.view(*proj_shape) + value_states = value_states.view(*proj_shape) + + src_len = key_states.size(1) + attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) + + if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): + raise ValueError( + f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is" + f" {attn_weights.size()}" + ) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, tgt_len, src_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" + ) + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + attn_weights = nn.functional.softmax(attn_weights, dim=-1) + + if layer_head_mask is not None: + if layer_head_mask.size() != (self.num_heads,): + raise ValueError( + f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" + f" {layer_head_mask.size()}" + ) + attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if output_attentions: + # this operation is a bit awkward, but it's required to + # make sure that attn_weights keeps its gradient. + # In order to do so, attn_weights have to be reshaped + # twice and have to be reused in the following + attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) + else: + attn_weights_reshaped = None + + attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) + + attn_output = torch.bmm(attn_probs, value_states) + + if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + + attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) + attn_output = attn_output.transpose(1, 2) + + # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be + # partitioned aross GPUs when using tensor-parallelism. + attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) + + attn_output = self.out_proj(attn_output) + + return attn_output, attn_weights_reshaped, past_key_value + + +class MvpEncoderLayer(nn.Module): + def __init__(self, config: MvpConfig): + super().__init__() + self.embed_dim = config.d_model + self.self_attn = MvpAttention( + embed_dim=self.embed_dim, + num_heads=config.encoder_attention_heads, + dropout=config.attention_dropout, + ) + self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) + self.dropout = config.dropout + self.activation_fn = ACT2FN[config.activation_function] + self.activation_dropout = config.activation_dropout + self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim) + self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim) + self.final_layer_norm = nn.LayerNorm(self.embed_dim) + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: torch.FloatTensor, + layer_head_mask: torch.FloatTensor, + self_attn_prompt: torch.FloatTensor, + output_attentions: Optional[bool] = False, + ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]: + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)` + attention_mask (`torch.FloatTensor`): attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size + `(encoder_attention_heads,)`. + self_attn_prompt (`torch.FloatTensor`): prompt of self attention of shape + `(2, encoder_attention_heads, pro_len, head_dim)`. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + """ + residual = hidden_states + hidden_states, attn_weights, _ = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + layer_head_mask=layer_head_mask, + attn_prompt=self_attn_prompt, + output_attentions=output_attentions, + ) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + hidden_states = self.self_attn_layer_norm(hidden_states) + + residual = hidden_states + hidden_states = self.activation_fn(self.fc1(hidden_states)) + hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) + hidden_states = self.fc2(hidden_states) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + hidden_states = self.final_layer_norm(hidden_states) + + if hidden_states.dtype == torch.float16 and ( + torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any() + ): + clamp_value = torch.finfo(hidden_states.dtype).max - 1000 + hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (attn_weights,) + + return outputs + + +class MvpDecoderLayer(nn.Module): + def __init__(self, config: MvpConfig): + super().__init__() + self.embed_dim = config.d_model + + self.self_attn = MvpAttention( + embed_dim=self.embed_dim, + num_heads=config.decoder_attention_heads, + dropout=config.attention_dropout, + is_decoder=True, + ) + self.dropout = config.dropout + self.activation_fn = ACT2FN[config.activation_function] + self.activation_dropout = config.activation_dropout + + self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) + self.encoder_attn = MvpAttention( + self.embed_dim, + config.decoder_attention_heads, + dropout=config.attention_dropout, + is_decoder=True, + ) + self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim) + self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim) + self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim) + self.final_layer_norm = nn.LayerNorm(self.embed_dim) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + layer_head_mask: Optional[torch.Tensor] = None, + cross_attn_layer_head_mask: Optional[torch.Tensor] = None, + self_attn_prompt: Optional[torch.Tensor] = None, + cross_attn_prompt: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = True, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`): attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + encoder_hidden_states (`torch.FloatTensor`): + cross attention input to the layer of shape `(batch, seq_len, embed_dim)` + encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size + `(encoder_attention_heads,)`. + cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of + size `(decoder_attention_heads,)`. + self_attn_prompt (`torch.FloatTensor`): prompt of self attention of shape + `(2, decoder_attention_heads, pro_len, head_dim)`. + cross_attn_prompt (`torch.FloatTensor`): prompt of cross attention of shape + `(2, decoder_attention_heads, pro_len, head_dim)`. + past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + """ + residual = hidden_states + + # Self Attention + # decoder uni-directional self-attention cached key/values tuple is at positions 1,2 + self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None + # add present self-attn cache to positions 1,2 of present_key_value tuple + hidden_states, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + past_key_value=self_attn_past_key_value, + attention_mask=attention_mask, + layer_head_mask=layer_head_mask, + attn_prompt=self_attn_prompt, + output_attentions=output_attentions, + ) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + hidden_states = self.self_attn_layer_norm(hidden_states) + + # Cross-Attention Block + cross_attn_present_key_value = None + cross_attn_weights = None + if encoder_hidden_states is not None: + residual = hidden_states + + # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple + cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None + hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn( + hidden_states=hidden_states, + key_value_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + layer_head_mask=cross_attn_layer_head_mask, + attn_prompt=cross_attn_prompt, + past_key_value=cross_attn_past_key_value, + output_attentions=output_attentions, + ) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + hidden_states = self.encoder_attn_layer_norm(hidden_states) + + # add cross-attn to positions 3,4 of present_key_value tuple + present_key_value = present_key_value + cross_attn_present_key_value + + # Fully Connected + residual = hidden_states + hidden_states = self.activation_fn(self.fc1(hidden_states)) + hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) + hidden_states = self.fc2(hidden_states) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + hidden_states = self.final_layer_norm(hidden_states) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights, cross_attn_weights) + + if use_cache: + outputs += (present_key_value,) + + return outputs + + +# Copied from transformers.models.bart.modeling_bart.BartClassificationHead with Bart->MVP +class MvpClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__( + self, + input_dim: int, + inner_dim: int, + num_classes: int, + pooler_dropout: float, + ): + super().__init__() + self.dense = nn.Linear(input_dim, inner_dim) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = nn.Linear(inner_dim, num_classes) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = self.dropout(hidden_states) + hidden_states = self.dense(hidden_states) + hidden_states = torch.tanh(hidden_states) + hidden_states = self.dropout(hidden_states) + hidden_states = self.out_proj(hidden_states) + return hidden_states + + +class MvpPrompt(nn.Module): + """Layer-wise prompt for encoder or decoder.""" + + def __init__(self, config, num_layers, num_heads): + super().__init__() + self.prompt_length = config.prompt_length + self.num_layers = num_layers + self.num_heads = num_heads + self.head_dim = config.d_model // num_heads + self.dropout = nn.Dropout(p=config.dropout) + self.prompt_embedding = nn.Embedding(config.prompt_length, config.d_model) + self.prompt_trans = nn.Sequential( + nn.Linear(config.d_model, config.prompt_mid_dim), + nn.GELU(), + nn.Linear(config.prompt_mid_dim, num_layers * 2 * config.d_model), + ) + + def forward(self, prompt_ids: torch.Tensor) -> Tuple[torch.Tensor]: + prompt = self.prompt_trans(self.prompt_embedding(prompt_ids)) + prompt = prompt.view(self.prompt_length, self.num_layers * 2, self.num_heads, self.head_dim) + prompt = self.dropout(prompt) + prompt = prompt.permute([1, 2, 0, 3]).split(2) + return prompt + + +class MvpPreTrainedModel(PreTrainedModel): + config_class = MvpConfig + base_model_prefix = "model" + supports_gradient_checkpointing = True + _keys_to_ignore_on_load_unexpected = [r"encoder.version", r"decoder.version"] + + def _init_weights(self, module): + std = self.config.init_std + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (MvpDecoder, MvpEncoder, MvpPrompt)): + module.gradient_checkpointing = value + + @property + def dummy_inputs(self): + pad_token = self.config.pad_token_id + input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device) + dummy_inputs = { + "attention_mask": input_ids.ne(pad_token), + "input_ids": input_ids, + } + return dummy_inputs + + +MVP_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`MvpConfig`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + +MVP_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`MvpTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): + Indices of decoder input sequence tokens in the vocabulary. + + Indices can be obtained using [`MvpTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are decoder input IDs?](../glossary#decoder-input-ids) + + Mvp uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` + is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). + + For translation and summarization training, `decoder_input_ids` should be provided. If no + `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right + for denoising pre-training following the paper. + decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): + Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also + be used by default. + + If you want to change padding behavior, you should read [`modeling_mvp._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0, + 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*): + Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`) + `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of + hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape + `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape + `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that + don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all + `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape + `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you + can choose to directly pass an embedded representation. This is useful if you want more control over how to + convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. + decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded + representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be + input (see `past_key_values`). This is useful if you want more control over how to convert + `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. + + If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value + of `inputs_embeds`. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + +MVP_CONDITIONAL_GENERATION_EXAMPLE = r""" + Example of summarization: + + Fine-tuning a model + ```python + >>> import torch + >>> from transformers import MvpTokenizer, MvpForConditionalGeneration + + >>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + >>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp") + + >>> inputs = tokenizer( + ... "Summarize: You may want to stick it to your boss and leave your job, but don't do it if these are your reasons.", + ... return_tensors="pt", + ... ) + >>> labels = tokenizer("Bad Reasons To Quit Your Job", return_tensors="pt")["input_ids"] + + >>> loss = model(**inputs, labels=labels).loss + >>> loss.backward() + ``` + + Inference after the model fine-tuned + ```python + >>> with torch.no_grad(): + ... generated_ids = model.generate(**inputs) + + >>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) + ``` +""" + +MVP_SEQUENCE_CLASSIFICATION_SAMPLE = r""" + Example of single-label classification: + + Fine-tuning a model on `num_labels` classes + ```python + >>> import torch + >>> from transformers import MvpTokenizer, MvpForSequenceClassification + + >>> num_labels = 2 # for example, this is a binary classification task + >>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + >>> model = MvpForSequenceClassification.from_pretrained("RUCAIBox/mvp", num_labels=num_labels) + + >>> inputs = tokenizer("Classify: Hello, my dog is cute", return_tensors="pt") + >>> labels = torch.tensor(1) # the real label for inputs + + >>> loss = model(**inputs, labels=labels).loss + >>> loss.backward() + ``` + + Inference after the model fine-tuned + ```python + >>> with torch.no_grad(): + ... logits = model(**inputs).logits + + >>> predicted_class_id = logits.argmax() + ``` +""" + +MVP_QUESTION_ANSWERING_SAMPLE = r""" + Example: + + Fine-tuning a model for extrative question answering, and our model also supports generative question answering + using `BartForConditionalGeneration` + ```python + >>> import torch + >>> from transformers import MvpTokenizer, MvpForQuestionAnswering + + >>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + >>> model = MvpForQuestionAnswering.from_pretrained("RUCAIBox/mvp") + + >>> inputs = tokenizer( + ... "Answer the following question: Who was Jim Henson? [SEP] Jim Henson was a nice puppet", + ... return_tensors="pt", + ... ) + >>> target_start_index = torch.tensor([18]) + >>> target_end_index = torch.tensor([19]) + + >>> loss = model(**inputs, start_positions=target_start_index, end_positions=target_end_index).loss + >>> loss.backward() + ``` + + Inference after the model fine-tuned + ```python + >>> with torch.no_grad(): + ... outputs = model(**inputs) + + >>> answer_start_index = outputs.start_logits.argmax() + >>> answer_end_index = outputs.end_logits.argmax() + + >>> predict_answer_tokens = inputs.input_ids[0, answer_start_index : answer_end_index + 1] + >>> predict_answer = tokenizer.decode(predict_answer_tokens) + ``` +""" + + +class MvpEncoder(MvpPreTrainedModel): + """ + Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a + [`MvpEncoderLayer`]. + + Args: + config: MvpConfig + embed_tokens (nn.Embedding): output embedding + use_prompt (bool): whether to use prompt + """ + + def __init__( + self, config: MvpConfig, embed_tokens: Optional[nn.Embedding] = None, use_prompt: Optional[bool] = False + ): + super().__init__(config) + + self.dropout = config.dropout + self.layerdrop = config.encoder_layerdrop + + embed_dim = config.d_model + self.padding_idx = config.pad_token_id + self.max_source_positions = config.max_position_embeddings + self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0 + + if embed_tokens is not None: + self.embed_tokens = embed_tokens + else: + self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx) + + self.embed_positions = MvpLearnedPositionalEmbedding( + config.max_position_embeddings, + embed_dim, + ) + self.layers = nn.ModuleList([MvpEncoderLayer(config) for _ in range(config.encoder_layers)]) + self.layernorm_embedding = nn.LayerNorm(embed_dim) + + self.use_prompt = use_prompt + if use_prompt: + self.prompt_length = config.prompt_length + self.self_attn_prompt = MvpPrompt( + config, + config.encoder_layers, + config.encoder_attention_heads, + ) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you + provide it. + + Indices can be obtained using [`MvpTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. + This is useful if you want more control over how to convert `input_ids` indices into associated vectors + than the model's internal embedding lookup matrix. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + else: + raise ValueError("You have to specify either input_ids or inputs_embeds") + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale + + embed_pos = self.embed_positions(input_shape) + + hidden_states = inputs_embeds + embed_pos + hidden_states = self.layernorm_embedding(hidden_states) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + + # layer-wise prompt + if self.use_prompt: + prompt_ids = torch.arange(self.prompt_length).to(self.device) + self_attn_prompt = self.self_attn_prompt(prompt_ids) + + # expand attention_mask + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype) + + encoder_states = () if output_hidden_states else None + all_attentions = () if output_attentions else None + + # check if head_mask has a correct number of layers specified if desired + if head_mask is not None: + if head_mask.size()[0] != (len(self.layers)): + raise ValueError( + f"The head_mask should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, encoder_layer in enumerate(self.layers): + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + dropout_probability = random.uniform(0, 1) + if self.training and (dropout_probability < self.layerdrop): # skip the layer + layer_outputs = (None, None) + else: + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs, output_attentions) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(encoder_layer), + hidden_states, + attention_mask, + (head_mask[idx] if head_mask is not None else None), + (self_attn_prompt[idx] if self.use_prompt else None), + ) + else: + layer_outputs = encoder_layer( + hidden_states, + attention_mask, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + self_attn_prompt=(self_attn_prompt[idx] if self.use_prompt else None), + output_attentions=output_attentions, + ) + + hidden_states = layer_outputs[0] + + if output_attentions: + all_attentions = all_attentions + (layer_outputs[1],) + + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + + if not return_dict: + return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) + return BaseModelOutput( + last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions + ) + + +class MvpDecoder(MvpPreTrainedModel): + """ + Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`MvpDecoderLayer`] + + Args: + config: MvpConfig + embed_tokens (nn.Embedding): output embedding + use_prompt (bool): whether to use prompt + """ + + def __init__( + self, config: MvpConfig, embed_tokens: Optional[nn.Embedding] = None, use_prompt: Optional[bool] = False + ): + super().__init__(config) + self.dropout = config.dropout + self.layerdrop = config.decoder_layerdrop + self.padding_idx = config.pad_token_id + self.max_target_positions = config.max_position_embeddings + self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 + + if embed_tokens is not None: + self.embed_tokens = embed_tokens + else: + self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx) + + self.embed_positions = MvpLearnedPositionalEmbedding( + config.max_position_embeddings, + config.d_model, + ) + self.layers = nn.ModuleList([MvpDecoderLayer(config) for _ in range(config.decoder_layers)]) + self.layernorm_embedding = nn.LayerNorm(config.d_model) + + self.use_prompt = use_prompt + if use_prompt: + self.prompt_length = config.prompt_length + self.self_attn_prompt = MvpPrompt( + config, + config.decoder_layers, + config.decoder_attention_heads, + ) + self.cross_attn_prompt = MvpPrompt( + config, + config.decoder_layers, + config.decoder_attention_heads, + ) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length): + # create causal mask + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + combined_attention_mask = None + if input_shape[-1] > 1: + combined_attention_mask = _make_causal_mask( + input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length + ).to(inputs_embeds.device) + + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]) + combined_attention_mask = ( + expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask + ) + + return combined_attention_mask + + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you + provide it. + + Indices can be obtained using [`MvpTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): + Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention + of the decoder. + encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): + Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values + selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing + cross-attention on hidden heads. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of + shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the + cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those + that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of + all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of + shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing + `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more + control over how to convert `input_ids` indices into associated vectors than the model's internal + embedding lookup matrix. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + # past_key_values_length + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale + + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, input_shape, inputs_embeds, past_key_values_length + ) + + # expand encoder attention mask + if encoder_hidden_states is not None and encoder_attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]) + + # embed positions + positions = self.embed_positions(input_shape, past_key_values_length) + + hidden_states = inputs_embeds + positions + hidden_states = self.layernorm_embedding(hidden_states) + + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + + # layer-wise prompt + if self.use_prompt: + prompt_ids = torch.arange(self.prompt_length).to(self.device) + self_attn_prompt = self.self_attn_prompt(prompt_ids) + cross_attn_prompt = self.cross_attn_prompt(prompt_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None + next_decoder_cache = () if use_cache else None + + # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired + for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]): + if attn_mask is not None: + if attn_mask.size()[0] != (len(self.layers)): + raise ValueError( + f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, decoder_layer in enumerate(self.layers): + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + if output_hidden_states: + all_hidden_states += (hidden_states,) + dropout_probability = random.uniform(0, 1) + if self.training and (dropout_probability < self.layerdrop): + continue + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + + if use_cache: + logger.warning( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, output_attentions, use_cache) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + head_mask[idx] if head_mask is not None else None, + cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None, + self_attn_prompt[idx] if self.use_prompt else None, + cross_attn_prompt[idx] if self.use_prompt else None, + None, + ) + else: + + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + cross_attn_layer_head_mask=( + cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None + ), + self_attn_prompt=(self_attn_prompt[idx] if self.use_prompt else None), + cross_attn_prompt=(cross_attn_prompt[idx] if self.use_prompt else None), + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[3 if output_attentions else 1],) + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + if encoder_hidden_states is not None: + all_cross_attentions += (layer_outputs[2],) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple( + v + for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions] + if v is not None + ) + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + cross_attentions=all_cross_attentions, + ) + + +@add_start_docstrings( + "The bare MVP Model outputting raw hidden-states without any specific head on top.", + MVP_START_DOCSTRING, +) +class MvpModel(MvpPreTrainedModel): + _keys_to_ignore_on_load_unexpected = [r"final_logits_bias", r"lm_head.weight"] + + def __init__(self, config: MvpConfig): + super().__init__(config) + + padding_idx, vocab_size = config.pad_token_id, config.vocab_size + self.use_prompt = config.use_prompt + self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx) + + self.encoder = MvpEncoder(config, self.shared, config.use_prompt) + self.decoder = MvpDecoder(config, self.shared, config.use_prompt) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.shared + + def set_input_embeddings(self, value): + self.shared = value + self.encoder.embed_tokens = self.shared + self.decoder.embed_tokens = self.shared + + def get_encoder(self): + return self.encoder + + def get_decoder(self): + return self.decoder + + def set_lightweight_tuning(self): + assert self.use_prompt, "If you want to use lightweight tuning, make sure that `use_prompt=True`." + + self.requires_grad_(False) + self.encoder.self_attn_prompt.requires_grad_(True) + self.decoder.self_attn_prompt.requires_grad_(True) + self.decoder.cross_attn_prompt.requires_grad_(True) + + @add_start_docstrings_to_model_forward(MVP_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + processor_class=_TOKENIZER_FOR_DOC, + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=Seq2SeqModelOutput, + config_class=_CONFIG_FOR_DOC, + expected_output=_EXPECTED_OUTPUT_SHAPE, + ) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_ids: Optional[torch.LongTensor] = None, + decoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + decoder_head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[List[torch.FloatTensor]] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + decoder_inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, Seq2SeqModelOutput]: + + # different to other models, Mvp automatically creates decoder_input_ids from + # input_ids if no decoder_input_ids are provided + if decoder_input_ids is None and decoder_inputs_embeds is None: + if input_ids is None: + raise ValueError( + "If no `decoder_input_ids` or `decoder_inputs_embeds` are " + "passed, `input_ids` cannot be `None`. Please pass either " + "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`." + ) + + decoder_input_ids = shift_tokens_right( + input_ids, self.config.pad_token_id, self.config.decoder_start_token_id + ) + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if encoder_outputs is None: + encoder_outputs = self.encoder( + input_ids=input_ids, + attention_mask=attention_mask, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True + elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): + encoder_outputs = BaseModelOutput( + last_hidden_state=encoder_outputs[0], + hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, + attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, + ) + + # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn) + decoder_outputs = self.decoder( + input_ids=decoder_input_ids, + attention_mask=decoder_attention_mask, + encoder_hidden_states=encoder_outputs[0], + encoder_attention_mask=attention_mask, + head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + inputs_embeds=decoder_inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + if not return_dict: + return decoder_outputs + encoder_outputs + + return Seq2SeqModelOutput( + last_hidden_state=decoder_outputs.last_hidden_state, + past_key_values=decoder_outputs.past_key_values, + decoder_hidden_states=decoder_outputs.hidden_states, + decoder_attentions=decoder_outputs.attentions, + cross_attentions=decoder_outputs.cross_attentions, + encoder_last_hidden_state=encoder_outputs.last_hidden_state, + encoder_hidden_states=encoder_outputs.hidden_states, + encoder_attentions=encoder_outputs.attentions, + ) + + +@add_start_docstrings( + "The MVP Model with a language modeling head. Can be used for various text generation tasks.", MVP_START_DOCSTRING +) +class MvpForConditionalGeneration(MvpPreTrainedModel): + def __init__(self, config: MvpConfig): + super().__init__(config) + self.model = MvpModel(config) + self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings))) + self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_encoder(self): + return self.model.get_encoder() + + def get_decoder(self): + return self.model.get_decoder() + + def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding: + new_embeddings = super().resize_token_embeddings(new_num_tokens) + self._resize_final_logits_bias(new_num_tokens) + return new_embeddings + + def _resize_final_logits_bias(self, new_num_tokens: int) -> None: + old_num_tokens = self.final_logits_bias.shape[-1] + if new_num_tokens <= old_num_tokens: + new_bias = self.final_logits_bias[:, :new_num_tokens] + else: + extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device) + new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1) + self.register_buffer("final_logits_bias", new_bias) + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_lightweight_tuning(self): + self.model.set_lightweight_tuning() + self.lm_head.requires_grad_(False) + + @add_start_docstrings_to_model_forward(MVP_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC) + @add_end_docstrings(MVP_CONDITIONAL_GENERATION_EXAMPLE) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_ids: Optional[torch.LongTensor] = None, + decoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + decoder_head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[List[torch.FloatTensor]] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + decoder_inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, Seq2SeqLMOutput]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Returns: + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if labels is not None: + if use_cache: + logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.") + use_cache = False + if decoder_input_ids is None and decoder_inputs_embeds is None: + decoder_input_ids = shift_tokens_right( + labels, self.config.pad_token_id, self.config.decoder_start_token_id + ) + + outputs = self.model( + input_ids, + attention_mask=attention_mask, + decoder_input_ids=decoder_input_ids, + encoder_outputs=encoder_outputs, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + decoder_inputs_embeds=decoder_inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias + + masked_lm_loss = None + if labels is not None: + loss_fct = CrossEntropyLoss() + masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1)) + + if not return_dict: + output = (lm_logits,) + outputs[1:] + return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output + + return Seq2SeqLMOutput( + loss=masked_lm_loss, + logits=lm_logits, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + def prepare_inputs_for_generation( + self, + decoder_input_ids, + past=None, + attention_mask=None, + head_mask=None, + decoder_head_mask=None, + cross_attn_head_mask=None, + use_cache=None, + encoder_outputs=None, + **kwargs + ): + # cut decoder_input_ids if past is used + if past is not None: + decoder_input_ids = decoder_input_ids[:, -1:] + + return { + "input_ids": None, # encoder_outputs is defined. input_ids not needed + "encoder_outputs": encoder_outputs, + "past_key_values": past, + "decoder_input_ids": decoder_input_ids, + "attention_mask": attention_mask, + "head_mask": head_mask, + "decoder_head_mask": decoder_head_mask, + "cross_attn_head_mask": cross_attn_head_mask, + "use_cache": use_cache, # change this to avoid caching (presumably for debugging) + } + + def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor): + return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id) + + @staticmethod + def _reorder_cache(past, beam_idx): + reordered_past = () + for layer_past in past: + # cached cross_attention states don't have to be reordered -> they are always the same + reordered_past += ( + tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:], + ) + return reordered_past + + +@add_start_docstrings( + """ + Mvp model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE + tasks. + """, + MVP_START_DOCSTRING, +) +class MvpForSequenceClassification(MvpPreTrainedModel): + _keys_to_ignore_on_load_unexpected = [r"final_logits_bias", r"lm_head.weight"] + + def __init__(self, config: MvpConfig, **kwargs): + super().__init__(config, **kwargs) + self.model = MvpModel(config) + self.classification_head = MvpClassificationHead( + config.d_model, + config.d_model, + config.num_labels, + config.classifier_dropout, + ) + + self.model._init_weights(self.classification_head.dense) + self.model._init_weights(self.classification_head.out_proj) + + def set_lightweight_tuning(self): + self.model.set_lightweight_tuning() + self.classification_head.requires_grad_(False) + + @add_start_docstrings_to_model_forward(MVP_INPUTS_DOCSTRING) + @add_end_docstrings(MVP_SEQUENCE_CLASSIFICATION_SAMPLE) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_ids: Optional[torch.LongTensor] = None, + decoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + decoder_head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + decoder_inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, Seq2SeqSequenceClassifierOutput]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + if labels is not None: + use_cache = False + + if input_ids is None and inputs_embeds is not None: + raise NotImplementedError( + f"Passing input embeddings is currently not supported for {self.__class__.__name__}" + ) + + outputs = self.model( + input_ids, + attention_mask=attention_mask, + decoder_input_ids=decoder_input_ids, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + encoder_outputs=encoder_outputs, + inputs_embeds=inputs_embeds, + decoder_inputs_embeds=decoder_inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = outputs[0] # last hidden state + + eos_mask = input_ids.eq(self.config.eos_token_id) + + if len(torch.unique_consecutive(eos_mask.sum(1))) > 1: + raise ValueError("All examples must have the same number of tokens.") + sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[ + :, -1, : + ] + logits = self.classification_head(sentence_representation) + + loss = None + if labels is not None: + if self.config.problem_type is None: + if self.config.num_labels == 1: + self.config.problem_type = "regression" + elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): + self.config.problem_type = "single_label_classification" + else: + self.config.problem_type = "multi_label_classification" + + if self.config.problem_type == "regression": + loss_fct = MSELoss() + if self.config.num_labels == 1: + loss = loss_fct(logits.squeeze(), labels.squeeze()) + else: + loss = loss_fct(logits, labels) + elif self.config.problem_type == "single_label_classification": + loss_fct = CrossEntropyLoss() + loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1)) + elif self.config.problem_type == "multi_label_classification": + loss_fct = BCEWithLogitsLoss() + loss = loss_fct(logits, labels) + if not return_dict: + output = (logits,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return Seq2SeqSequenceClassifierOutput( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + +@add_start_docstrings( + """ + MVP Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layer + on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + MVP_START_DOCSTRING, +) +class MvpForQuestionAnswering(MvpPreTrainedModel): + _keys_to_ignore_on_load_unexpected = [r"final_logits_bias", r"lm_head.weight"] + + def __init__(self, config): + super().__init__(config) + + config.num_labels = 2 + self.num_labels = config.num_labels + + self.model = MvpModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels) + + self.model._init_weights(self.qa_outputs) + + def set_lightweight_tuning(self): + self.model.set_lightweight_tuning() + self.qa_outputs.requires_grad_(False) + + @add_start_docstrings_to_model_forward(MVP_INPUTS_DOCSTRING) + @add_end_docstrings(MVP_QUESTION_ANSWERING_SAMPLE) + def forward( + self, + input_ids: torch.Tensor = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_ids: Optional[torch.LongTensor] = None, + decoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + decoder_head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[List[torch.FloatTensor]] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + decoder_inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, Seq2SeqQuestionAnsweringModelOutput]: + r""" + start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the start of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence + are not taken into account for computing the loss. + end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the end of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence + are not taken into account for computing the loss. + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + if start_positions is not None and end_positions is not None: + use_cache = False + + outputs = self.model( + input_ids, + attention_mask=attention_mask, + decoder_input_ids=decoder_input_ids, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + encoder_outputs=encoder_outputs, + inputs_embeds=inputs_embeds, + decoder_inputs_embeds=decoder_inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + sequence_output = outputs[0] + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + total_loss = None + if start_positions is not None and end_positions is not None: + # If we are on multi-GPU, split add a dimension + if len(start_positions.size()) > 1: + start_positions = start_positions.squeeze(-1) + if len(end_positions.size()) > 1: + end_positions = end_positions.squeeze(-1) + # sometimes the start/end positions are outside our model inputs, we ignore these terms + ignored_index = start_logits.size(1) + start_positions = start_positions.clamp(0, ignored_index) + end_positions = end_positions.clamp(0, ignored_index) + + loss_fct = CrossEntropyLoss(ignore_index=ignored_index) + start_loss = loss_fct(start_logits, start_positions) + end_loss = loss_fct(end_logits, end_positions) + total_loss = (start_loss + end_loss) / 2 + + if not return_dict: + output = ( + start_logits, + end_logits, + ) + outputs[1:] + return ((total_loss,) + output) if total_loss is not None else output + + return Seq2SeqQuestionAnsweringModelOutput( + loss=total_loss, + start_logits=start_logits, + end_logits=end_logits, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + +# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->Mvp +class MvpDecoderWrapper(MvpPreTrainedModel): + """ + This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is + used in combination with the [`EncoderDecoderModel`] framework. + """ + + def __init__(self, config): + super().__init__(config) + self.decoder = MvpDecoder(config) + + def forward(self, *args, **kwargs): + return self.decoder(*args, **kwargs) + + +class MvpForCausalLM(MvpPreTrainedModel): + def __init__(self, config): + config = copy.deepcopy(config) + config.is_decoder = True + config.is_encoder_decoder = False + super().__init__(config) + self.model = MvpDecoderWrapper(config) + + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.decoder.embed_tokens + + def set_input_embeddings(self, value): + self.model.decoder.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model.decoder = decoder + + def get_decoder(self): + return self.model.decoder + + def set_lightweight_tuning(self): + self.model.set_lightweight_tuning() + self.lm_head.requires_grad_(False) + + @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]: + r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you + provide it. + + Indices can be obtained using [`MvpTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention + if the model is configured as a decoder. + encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used + in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: + head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of + shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional + tensors are only required when the model is used as a decoder in a Sequence to Sequence model. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the + cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those + that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of + all `decoder_input_ids` of shape `(batch_size, sequence_length)`. + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + + Returns: + + Example: + + ```python + >>> from transformers import MvpTokenizer, MvpForCausalLM + + >>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + >>> model = MvpForCausalLM.from_pretrained("RUCAIBox/mvp", add_cross_attention=False) + + >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt") + >>> outputs = model(**inputs) + + >>> logits = outputs.logits + >>> list(logits.shape) + [1, 8, 50267] + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs = self.model.decoder( + input_ids=input_ids, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + head_mask=head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + logits = self.lm_head(outputs[0]) + + loss = None + if labels is not None: + loss_fct = CrossEntropyLoss() + loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1)) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithCrossAttentions( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + cross_attentions=outputs.cross_attentions, + ) + + def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, use_cache=None, **kwargs): + # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly + if attention_mask is None: + attention_mask = input_ids.new_ones(input_ids.shape) + + if past: + input_ids = input_ids[:, -1:] + # first step, decoder_cached_states are empty + return { + "input_ids": input_ids, # encoder_outputs is defined. input_ids not needed + "attention_mask": attention_mask, + "past_key_values": past, + "use_cache": use_cache, + } + + @staticmethod + def _reorder_cache(past, beam_idx): + reordered_past = () + for layer_past in past: + reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),) + return reordered_past diff --git a/src/transformers/models/mvp/tokenization_mvp.py b/src/transformers/models/mvp/tokenization_mvp.py new file mode 100644 index 0000000000..3d5d606d63 --- /dev/null +++ b/src/transformers/models/mvp/tokenization_mvp.py @@ -0,0 +1,404 @@ +# coding=utf-8 +# Copyright 2022 The Facebook AI Research Team Authors and The HuggingFace Inc. team. +# +# 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. + +import json +import os +from functools import lru_cache +from typing import List, Optional, Tuple + +import regex as re + +from ...tokenization_utils import AddedToken, PreTrainedTokenizer +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt"} + +# See all MVP models at https://huggingface.co/models?filter=mvp +PRETRAINED_VOCAB_FILES_MAP = { + "vocab_file": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/vocab.json", + }, + "added_tokens.json": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/added_tokens.json", + }, + "merges_file": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/merges.txt", + }, +} + +PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { + "RUCAIBox/mvp": 1024, +} + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control + characters the bpe code barfs on. + + The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab + if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for + decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup + tables between utf-8 bytes and unicode strings. + """ + bs = ( + list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1)) + ) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8 + n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +def get_pairs(word): + """ + Return set of symbol pairs in a word. + + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + + +class MvpTokenizer(PreTrainedTokenizer): + """ + Constructs a MVP tokenizer, which is smilar to the RoBERTa tokenizer, using byte-level Byte-Pair-Encoding. + + This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will + be encoded differently whether it is at the beginning of the sentence (without space) or not: + + ``` + >>> from transformers import MvpTokenizer + >>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + >>> tokenizer("Hello world")['input_ids'] + [0, 31414, 232, 2] + >>> tokenizer(" Hello world")['input_ids'] + [0, 20920, 232, 2] + ``` + + You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you + call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance. + + + + When used with `is_split_into_words=True`, this tokenizer will add a space before each word (even the first one). + + + + This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to + this superclass for more information regarding those methods. + + Args: + vocab_file (`str`): + Path to the vocabulary file. + merges_file (`str`): + Path to the merges file. + errors (`str`, *optional*, defaults to `"replace"`): + Paradigm to follow when decoding bytes to UTF-8. See + [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information. + bos_token (`str`, *optional*, defaults to `""`): + The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token. + + + + When building a sequence using special tokens, this is not the token that is used for the beginning of + sequence. The token used is the `cls_token`. + + + + eos_token (`str`, *optional*, defaults to `""`): + The end of sequence token. + + + + When building a sequence using special tokens, this is not the token that is used for the end of sequence. + The token used is the `sep_token`. + + + + sep_token (`str`, *optional*, defaults to `""`): + The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for + sequence classification or for a text and a question for question answering. It is also used as the last + token of a sequence built with special tokens. + cls_token (`str`, *optional*, defaults to `""`): + The classifier token which is used when doing sequence classification (classification of the whole sequence + instead of per-token classification). It is the first token of the sequence when built with special tokens. + unk_token (`str`, *optional*, defaults to `""`): + The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this + token instead. + pad_token (`str`, *optional*, defaults to `""`): + The token used for padding, for example when batching sequences of different lengths. + mask_token (`str`, *optional*, defaults to `""`): + The token used for masking values. This is the token used when training this model with masked language + modeling. This is the token which the model will try to predict. + add_prefix_space (`bool`, *optional*, defaults to `False`): + Whether or not to add an initial space to the input. This allows to treat the leading word just as any + other word. (MVP tokenizer detect beginning of words by the preceding space). + """ + + vocab_files_names = VOCAB_FILES_NAMES + pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP + max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES + model_input_names = ["input_ids", "attention_mask"] + + def __init__( + self, + vocab_file, + merges_file, + errors="replace", + bos_token="", + eos_token="", + sep_token="", + cls_token="", + unk_token="", + pad_token="", + mask_token="", + add_prefix_space=False, + **kwargs + ): + bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token + eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token + sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token + cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token + unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token + pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token + + # Mask token behave like a normal word, i.e. include the space before it + mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token + + super().__init__( + errors=errors, + bos_token=bos_token, + eos_token=eos_token, + unk_token=unk_token, + sep_token=sep_token, + cls_token=cls_token, + pad_token=pad_token, + mask_token=mask_token, + add_prefix_space=add_prefix_space, + **kwargs, + ) + with open(vocab_file, encoding="utf-8") as vocab_handle: + self.encoder = json.load(vocab_handle) + self.decoder = {v: k for k, v in self.encoder.items()} + self.errors = errors # how to handle errors in decoding + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + with open(merges_file, encoding="utf-8") as merges_handle: + bpe_merges = merges_handle.read().split("\n")[1:-1] + bpe_merges = [tuple(merge.split()) for merge in bpe_merges] + self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges)))) + self.cache = {} + self.add_prefix_space = add_prefix_space + + # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions + self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""") + + @property + def vocab_size(self): + return len(self.encoder) + + def get_vocab(self): + return dict(self.encoder, **self.added_tokens_encoder) + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token) + pairs = get_pairs(word) + + if not pairs: + return token + + while True: + bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf"))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + except ValueError: + new_word.extend(word[i:]) + break + else: + new_word.extend(word[i:j]) + i = j + + if word[i] == first and i < len(word) - 1 and word[i + 1] == second: + new_word.append(first + second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = " ".join(word) + self.cache[token] = word + return word + + def _tokenize(self, text): + """Tokenize a string.""" + bpe_tokens = [] + for token in re.findall(self.pat, text): + token = "".join( + self.byte_encoder[b] for b in token.encode("utf-8") + ) # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case) + bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" ")) + return bpe_tokens + + def _convert_token_to_id(self, token): + """Converts a token (str) in an id using the vocab.""" + return self.encoder.get(token, self.encoder.get(self.unk_token)) + + def _convert_id_to_token(self, index): + """Converts an index (integer) in a token (str) using the vocab.""" + return self.decoder.get(index) + + def convert_tokens_to_string(self, tokens): + """Converts a sequence of tokens (string) in a single string.""" + text = "".join(tokens) + text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors) + return text + + def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]: + if not os.path.isdir(save_directory): + logger.error(f"Vocabulary path ({save_directory}) should be a directory") + return + vocab_file = os.path.join( + save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"] + ) + merge_file = os.path.join( + save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"] + ) + + with open(vocab_file, "w", encoding="utf-8") as f: + f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n") + + index = 0 + with open(merge_file, "w", encoding="utf-8") as writer: + writer.write("#version: 0.2\n") + for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]): + if index != token_index: + logger.warning( + f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive." + " Please check that the tokenizer is not corrupted!" + ) + index = token_index + writer.write(" ".join(bpe_tokens) + "\n") + index += 1 + + return vocab_file, merge_file + + def build_inputs_with_special_tokens( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None + ) -> List[int]: + """ + Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and + adding special tokens. A MVP sequence has the following format: + + - single sequence: ` X ` + - pair of sequences: ` A B ` + + Args: + token_ids_0 (`List[int]`): + List of IDs to which the special tokens will be added. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + + Returns: + `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens. + """ + if token_ids_1 is None: + return [self.cls_token_id] + token_ids_0 + [self.sep_token_id] + cls = [self.cls_token_id] + sep = [self.sep_token_id] + return cls + token_ids_0 + sep + sep + token_ids_1 + sep + + def get_special_tokens_mask( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False + ) -> List[int]: + """ + Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding + special tokens using the tokenizer `prepare_for_model` method. + + Args: + token_ids_0 (`List[int]`): + List of IDs. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + already_has_special_tokens (`bool`, *optional*, defaults to `False`): + Whether or not the token list is already formatted with special tokens for the model. + + Returns: + `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token. + """ + if already_has_special_tokens: + return super().get_special_tokens_mask( + token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True + ) + + if token_ids_1 is None: + return [1] + ([0] * len(token_ids_0)) + [1] + return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1] + + def create_token_type_ids_from_sequences( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None + ) -> List[int]: + """ + Create a mask from the two sequences passed to be used in a sequence-pair classification task. MVP does not + make use of token type ids, therefore a list of zeros is returned. + + Args: + token_ids_0 (`List[int]`): + List of IDs. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + + Returns: + `List[int]`: List of zeros. + """ + sep = [self.sep_token_id] + cls = [self.cls_token_id] + + if token_ids_1 is None: + return len(cls + token_ids_0 + sep) * [0] + return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0] + + def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs): + add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space) + if (is_split_into_words or add_prefix_space) and (len(text) > 0 and not text[0].isspace()): + text = " " + text + return (text, kwargs) diff --git a/src/transformers/models/mvp/tokenization_mvp_fast.py b/src/transformers/models/mvp/tokenization_mvp_fast.py new file mode 100644 index 0000000000..00b7a5c665 --- /dev/null +++ b/src/transformers/models/mvp/tokenization_mvp_fast.py @@ -0,0 +1,287 @@ +# coding=utf-8 +# Copyright 2022 The Facebook AI Research Team Authors and The HuggingFace Inc. team. +# +# 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. + +import json +from typing import List, Optional, Tuple + +from tokenizers import pre_tokenizers, processors + +from ...tokenization_utils_base import AddedToken, BatchEncoding +from ...tokenization_utils_fast import PreTrainedTokenizerFast +from ...utils import logging +from .tokenization_mvp import MvpTokenizer + + +logger = logging.get_logger(__name__) + + +VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"} + +# See all MVP models at https://huggingface.co/models?filter=mvp +PRETRAINED_VOCAB_FILES_MAP = { + "vocab_file": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/vocab.json", + }, + "added_tokens.json": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/added_tokens.json", + }, + "merges_file": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/merges.txt", + }, + "tokenizer_file": { + "RUCAIBox/mvp": "https://huggingface.co/RUCAIBox/mvp/resolve/main/tokenizer.json", + }, +} + +PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { + "RUCAIBox/mvp": 1024, +} + + +class MvpTokenizerFast(PreTrainedTokenizerFast): + r""" + Construct a "fast" MVP tokenizer (backed by HuggingFace's *tokenizers* library), derived from the GPT-2 tokenizer, + using byte-level Byte-Pair-Encoding. + + This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will + be encoded differently whether it is at the beginning of the sentence (without space) or not: + + ``` + >>> from transformers import MvpTokenizerFast + >>> tokenizer = MvpTokenizerFast.from_pretrained("RUCAIBox/mvp") + >>> tokenizer("Hello world")['input_ids'] + [0, 31414, 232, 2] + >>> tokenizer(" Hello world")['input_ids'] + [0, 20920, 232, 2] + ``` + + You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you + call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance. + + + + When used with `is_split_into_words=True`, this tokenizer needs to be instantiated with `add_prefix_space=True`. + + + + This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should + refer to this superclass for more information regarding those methods. + + Args: + vocab_file (`str`): + Path to the vocabulary file. + merges_file (`str`): + Path to the merges file. + errors (`str`, *optional*, defaults to `"replace"`): + Paradigm to follow when decoding bytes to UTF-8. See + [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information. + bos_token (`str`, *optional*, defaults to `""`): + The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token. + + + + When building a sequence using special tokens, this is not the token that is used for the beginning of + sequence. The token used is the `cls_token`. + + + + eos_token (`str`, *optional*, defaults to `""`): + The end of sequence token. + + + + When building a sequence using special tokens, this is not the token that is used for the end of sequence. + The token used is the `sep_token`. + + + + sep_token (`str`, *optional*, defaults to `""`): + The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for + sequence classification or for a text and a question for question answering. It is also used as the last + token of a sequence built with special tokens. + cls_token (`str`, *optional*, defaults to `""`): + The classifier token which is used when doing sequence classification (classification of the whole sequence + instead of per-token classification). It is the first token of the sequence when built with special tokens. + unk_token (`str`, *optional*, defaults to `""`): + The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this + token instead. + pad_token (`str`, *optional*, defaults to `""`): + The token used for padding, for example when batching sequences of different lengths. + mask_token (`str`, *optional*, defaults to `""`): + The token used for masking values. This is the token used when training this model with masked language + modeling. This is the token which the model will try to predict. + add_prefix_space (`bool`, *optional*, defaults to `False`): + Whether or not to add an initial space to the input. This allows to treat the leading word just as any + other word. (MVP tokenizer detect beginning of words by the preceding space). + trim_offsets (`bool`, *optional*, defaults to `True`): + Whether the post processing step should trim offsets to avoid including whitespaces. + """ + vocab_files_names = VOCAB_FILES_NAMES + pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP + max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES + model_input_names = ["input_ids", "attention_mask"] + slow_tokenizer_class = MvpTokenizer + + def __init__( + self, + vocab_file=None, + merges_file=None, + tokenizer_file=None, + errors="replace", + bos_token="", + eos_token="", + sep_token="", + cls_token="", + unk_token="", + pad_token="", + mask_token="", + add_prefix_space=False, + trim_offsets=True, + **kwargs + ): + super().__init__( + vocab_file, + merges_file, + tokenizer_file=tokenizer_file, + errors=errors, + bos_token=bos_token, + eos_token=eos_token, + sep_token=sep_token, + cls_token=cls_token, + unk_token=unk_token, + pad_token=pad_token, + mask_token=mask_token, + add_prefix_space=add_prefix_space, + trim_offsets=trim_offsets, + **kwargs, + ) + + pre_tok_state = json.loads(self.backend_tokenizer.pre_tokenizer.__getstate__()) + if pre_tok_state.get("add_prefix_space", add_prefix_space) != add_prefix_space: + pre_tok_class = getattr(pre_tokenizers, pre_tok_state.pop("type")) + pre_tok_state["add_prefix_space"] = add_prefix_space + self.backend_tokenizer.pre_tokenizer = pre_tok_class(**pre_tok_state) + + self.add_prefix_space = add_prefix_space + + # the pre_tokenizer is already updated in the GPT2TokenizerFast `__init__` + tokenizer_component = "post_processor" + tokenizer_component_instance = getattr(self.backend_tokenizer, tokenizer_component, None) + if tokenizer_component_instance: + state = json.loads(tokenizer_component_instance.__getstate__()) + + # The lists 'sep' and 'cls' must be cased in tuples for the object `post_processor_class` + if "sep" in state: + state["sep"] = tuple(state["sep"]) + if "cls" in state: + state["cls"] = tuple(state["cls"]) + + changes_to_apply = False + + if state.get("add_prefix_space", add_prefix_space) != add_prefix_space: + state["add_prefix_space"] = add_prefix_space + changes_to_apply = True + + if state.get("trim_offsets", trim_offsets) != trim_offsets: + state["trim_offsets"] = trim_offsets + changes_to_apply = True + + if changes_to_apply: + component_class = getattr(processors, state.pop("type")) + new_value = component_class(**state) + setattr(self.backend_tokenizer, tokenizer_component, new_value) + + @property + def mask_token(self) -> str: + """ + `str`: Mask token, to use when training a model with masked-language modeling. Log an error if used while not + having been set. + + MVP tokenizer has a special mask token to be usable in the fill-mask pipeline. The mask token will greedily + comprise the space before the **. + """ + if self._mask_token is None: + if self.verbose: + logger.error("Using mask_token, but it is not set yet.") + return None + return str(self._mask_token) + + @mask_token.setter + def mask_token(self, value): + """ + Overriding the default behavior of the mask token to have it eat the space before it. + + This is needed to preserve backward compatibility with all the previously used models based on Mvp. + """ + # Mask token behave like a normal word, i.e. include the space before it + # So we set lstrip to True + value = AddedToken(value, lstrip=True, rstrip=False) if isinstance(value, str) else value + self._mask_token = value + + def _batch_encode_plus(self, *args, **kwargs) -> BatchEncoding: + is_split_into_words = kwargs.get("is_split_into_words", False) + + if is_split_into_words and not self.add_prefix_space: + raise ValueError( + f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True " + "to use it with pretokenized inputs." + ) + + return super()._batch_encode_plus(*args, **kwargs) + + def _encode_plus(self, *args, **kwargs) -> BatchEncoding: + is_split_into_words = kwargs.get("is_split_into_words", False) + + if is_split_into_words and not self.add_prefix_space: + raise ValueError( + f"You need to instantiate {self.__class__.__name__} with add_prefix_space=True " + "to use it with pretokenized inputs." + ) + + return super()._encode_plus(*args, **kwargs) + + def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]: + files = self._tokenizer.model.save(save_directory, name=filename_prefix) + return tuple(files) + + def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None): + output = [self.bos_token_id] + token_ids_0 + [self.eos_token_id] + if token_ids_1 is None: + return output + + return output + [self.eos_token_id] + token_ids_1 + [self.eos_token_id] + + def create_token_type_ids_from_sequences( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None + ) -> List[int]: + """ + Create a mask from the two sequences passed to be used in a sequence-pair classification task. MVP does not + make use of token type ids, therefore a list of zeros is returned. + + Args: + token_ids_0 (`List[int]`): + List of IDs. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + + Returns: + `List[int]`: List of zeros. + """ + sep = [self.sep_token_id] + cls = [self.cls_token_id] + + if token_ids_1 is None: + return len(cls + token_ids_0 + sep) * [0] + return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0] diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index e8d6ae82d3..4a9279600c 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -3185,6 +3185,51 @@ class MT5Model(metaclass=DummyObject): requires_backends(self, ["torch"]) +MVP_PRETRAINED_MODEL_ARCHIVE_LIST = None + + +class MvpForCausalLM(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class MvpForConditionalGeneration(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class MvpForQuestionAnswering(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class MvpForSequenceClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class MvpModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class MvpPreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + NEZHA_PRETRAINED_MODEL_ARCHIVE_LIST = None diff --git a/src/transformers/utils/dummy_tokenizers_objects.py b/src/transformers/utils/dummy_tokenizers_objects.py index d6a63b633a..80f6e9c42e 100644 --- a/src/transformers/utils/dummy_tokenizers_objects.py +++ b/src/transformers/utils/dummy_tokenizers_objects.py @@ -262,6 +262,13 @@ class MT5TokenizerFast(metaclass=DummyObject): requires_backends(self, ["tokenizers"]) +class MvpTokenizerFast(metaclass=DummyObject): + _backends = ["tokenizers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["tokenizers"]) + + class OpenAIGPTTokenizerFast(metaclass=DummyObject): _backends = ["tokenizers"] diff --git a/tests/models/mvp/__init__.py b/tests/models/mvp/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/models/mvp/test_modeling_mvp.py b/tests/models/mvp/test_modeling_mvp.py new file mode 100644 index 0000000000..e0247d4233 --- /dev/null +++ b/tests/models/mvp/test_modeling_mvp.py @@ -0,0 +1,787 @@ +# coding=utf-8 +# Copyright 2021, The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" Testing suite for the PyTorch MVP model. """ + + +import copy +import tempfile +import unittest + +import timeout_decorator # noqa + +from transformers import MvpConfig, is_torch_available +from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device +from transformers.utils import cached_property + +from ...generation.test_generation_utils import GenerationTesterMixin +from ...test_configuration_common import ConfigTester +from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor + + +if is_torch_available(): + import torch + + from transformers import ( + MvpForCausalLM, + MvpForConditionalGeneration, + MvpForQuestionAnswering, + MvpForSequenceClassification, + MvpModel, + MvpTokenizer, + ) + from transformers.models.mvp.modeling_mvp import MvpDecoder, MvpEncoder, shift_tokens_right + + +def prepare_mvp_inputs_dict( + config, + input_ids, + decoder_input_ids=None, + attention_mask=None, + decoder_attention_mask=None, + head_mask=None, + decoder_head_mask=None, + cross_attn_head_mask=None, +): + if attention_mask is None: + attention_mask = input_ids.ne(config.pad_token_id) + if decoder_attention_mask is None: + decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id) + if head_mask is None: + head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device) + if decoder_head_mask is None: + decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device) + if cross_attn_head_mask is None: + cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device) + return { + "input_ids": input_ids, + "decoder_input_ids": decoder_input_ids, + "attention_mask": attention_mask, + "decoder_attention_mask": attention_mask, + "head_mask": head_mask, + "decoder_head_mask": decoder_head_mask, + "cross_attn_head_mask": cross_attn_head_mask, + } + + +class MvpModelTester: + def __init__( + self, + parent, + batch_size=13, + seq_length=7, + is_training=True, + use_labels=False, + vocab_size=99, + hidden_size=16, + num_hidden_layers=2, + num_attention_heads=4, + intermediate_size=4, + hidden_act="gelu", + hidden_dropout_prob=0.1, + attention_probs_dropout_prob=0.1, + max_position_embeddings=20, + eos_token_id=2, + pad_token_id=1, + bos_token_id=0, + ): + self.parent = parent + self.batch_size = batch_size + self.seq_length = seq_length + self.is_training = is_training + self.use_labels = use_labels + self.vocab_size = vocab_size + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.intermediate_size = intermediate_size + self.hidden_act = hidden_act + self.hidden_dropout_prob = hidden_dropout_prob + self.attention_probs_dropout_prob = attention_probs_dropout_prob + self.max_position_embeddings = max_position_embeddings + self.eos_token_id = eos_token_id + self.pad_token_id = pad_token_id + self.bos_token_id = bos_token_id + + def prepare_config_and_inputs(self): + input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) + input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp( + 3, + ) + input_ids[:, -1] = self.eos_token_id # Eos Token + + decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) + + config = self.get_config() + inputs_dict = prepare_mvp_inputs_dict(config, input_ids, decoder_input_ids) + return config, inputs_dict + + def get_config(self): + return MvpConfig( + vocab_size=self.vocab_size, + d_model=self.hidden_size, + encoder_layers=self.num_hidden_layers, + decoder_layers=self.num_hidden_layers, + encoder_attention_heads=self.num_attention_heads, + decoder_attention_heads=self.num_attention_heads, + encoder_ffn_dim=self.intermediate_size, + decoder_ffn_dim=self.intermediate_size, + dropout=self.hidden_dropout_prob, + attention_dropout=self.attention_probs_dropout_prob, + max_position_embeddings=self.max_position_embeddings, + eos_token_id=self.eos_token_id, + bos_token_id=self.bos_token_id, + pad_token_id=self.pad_token_id, + ) + + def get_pipeline_config(self): + config = self.get_config() + config.max_position_embeddings = 100 + config.vocab_size = 300 + return config + + def prepare_config_and_inputs_for_common(self): + config, inputs_dict = self.prepare_config_and_inputs() + return config, inputs_dict + + def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict): + model = MvpModel(config=config).get_decoder().to(torch_device).eval() + input_ids = inputs_dict["input_ids"] + attention_mask = inputs_dict["attention_mask"] + head_mask = inputs_dict["head_mask"] + + # first forward pass + outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True) + + output, past_key_values = outputs.to_tuple() + + # create hypothetical multiple next token and extent to next_input_ids + next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size) + next_attn_mask = ids_tensor((self.batch_size, 3), 2) + + # append to next input_ids and + next_input_ids = torch.cat([input_ids, next_tokens], dim=-1) + next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1) + + output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"] + output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[ + "last_hidden_state" + ] + + # select random slice + random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() + output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach() + output_from_past_slice = output_from_past[:, :, random_slice_idx].detach() + + self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1]) + + # test that outputs are equal for slice + self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)) + + def check_encoder_decoder_model_standalone(self, config, inputs_dict): + model = MvpModel(config=config).to(torch_device).eval() + outputs = model(**inputs_dict) + + encoder_last_hidden_state = outputs.encoder_last_hidden_state + last_hidden_state = outputs.last_hidden_state + + with tempfile.TemporaryDirectory() as tmpdirname: + encoder = model.get_encoder() + encoder.save_pretrained(tmpdirname) + encoder = MvpEncoder.from_pretrained(tmpdirname).to(torch_device) + + encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[ + 0 + ] + + self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3) + + with tempfile.TemporaryDirectory() as tmpdirname: + decoder = model.get_decoder() + decoder.save_pretrained(tmpdirname) + decoder = MvpDecoder.from_pretrained(tmpdirname).to(torch_device) + + last_hidden_state_2 = decoder( + input_ids=inputs_dict["decoder_input_ids"], + attention_mask=inputs_dict["decoder_attention_mask"], + encoder_hidden_states=encoder_last_hidden_state, + encoder_attention_mask=inputs_dict["attention_mask"], + )[0] + + self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3) + + +@require_torch +class MvpHeadTests(unittest.TestCase): + vocab_size = 99 + + def _get_config_and_data(self): + input_ids = torch.tensor( + [ + [71, 82, 18, 33, 46, 91, 2], + [68, 34, 26, 58, 30, 82, 2], + [5, 97, 17, 39, 94, 40, 2], + [76, 83, 94, 25, 70, 78, 2], + [87, 59, 41, 35, 48, 66, 2], + [55, 13, 16, 58, 5, 2, 1], # note padding + [64, 27, 31, 51, 12, 75, 2], + [52, 64, 86, 17, 83, 39, 2], + [48, 61, 9, 24, 71, 82, 2], + [26, 1, 60, 48, 22, 13, 2], + [21, 5, 62, 28, 14, 76, 2], + [45, 98, 37, 86, 59, 48, 2], + [70, 70, 50, 9, 28, 0, 2], + ], + dtype=torch.long, + device=torch_device, + ) + + batch_size = input_ids.shape[0] + config = MvpConfig( + vocab_size=self.vocab_size, + d_model=24, + encoder_layers=2, + decoder_layers=2, + encoder_attention_heads=2, + decoder_attention_heads=2, + encoder_ffn_dim=32, + decoder_ffn_dim=32, + max_position_embeddings=48, + eos_token_id=2, + pad_token_id=1, + bos_token_id=0, + ) + return config, input_ids, batch_size + + def test_sequence_classification_forward(self): + config, input_ids, batch_size = self._get_config_and_data() + labels = _long_tensor([2] * batch_size).to(torch_device) + config.num_labels = 3 + model = MvpForSequenceClassification(config) + model.to(torch_device) + outputs = model(input_ids=input_ids, decoder_input_ids=input_ids, labels=labels) + expected_shape = torch.Size((batch_size, config.num_labels)) + self.assertEqual(outputs["logits"].shape, expected_shape) + self.assertIsInstance(outputs["loss"].item(), float) + + def test_question_answering_forward(self): + config, input_ids, batch_size = self._get_config_and_data() + sequence_labels = ids_tensor([batch_size], 2).to(torch_device) + model = MvpForQuestionAnswering(config) + model.to(torch_device) + outputs = model( + input_ids=input_ids, + start_positions=sequence_labels, + end_positions=sequence_labels, + ) + + self.assertEqual(outputs["start_logits"].shape, input_ids.shape) + self.assertEqual(outputs["end_logits"].shape, input_ids.shape) + self.assertIsInstance(outputs["loss"].item(), float) + + @timeout_decorator.timeout(1) + def test_lm_forward(self): + config, input_ids, batch_size = self._get_config_and_data() + lm_labels = ids_tensor([batch_size, input_ids.shape[1]], self.vocab_size).to(torch_device) + lm_model = MvpForConditionalGeneration(config) + lm_model.to(torch_device) + outputs = lm_model(input_ids=input_ids, labels=lm_labels) + expected_shape = (batch_size, input_ids.shape[1], config.vocab_size) + self.assertEqual(outputs["logits"].shape, expected_shape) + self.assertIsInstance(outputs["loss"].item(), float) + + def test_lm_uneven_forward(self): + config = MvpConfig( + vocab_size=self.vocab_size, + d_model=14, + encoder_layers=2, + decoder_layers=2, + encoder_attention_heads=2, + decoder_attention_heads=2, + encoder_ffn_dim=8, + decoder_ffn_dim=8, + max_position_embeddings=48, + ) + lm_model = MvpForConditionalGeneration(config).to(torch_device) + context = torch.tensor( + [[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], device=torch_device, dtype=torch.long + ) + summary = torch.tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], device=torch_device, dtype=torch.long) + outputs = lm_model(input_ids=context, decoder_input_ids=summary, labels=summary) + expected_shape = (*summary.shape, config.vocab_size) + self.assertEqual(outputs["logits"].shape, expected_shape) + + def test_generate_beam_search(self): + input_ids = torch.tensor([[71, 82, 2], [68, 34, 2]], device=torch_device, dtype=torch.long) + config = MvpConfig( + vocab_size=self.vocab_size, + d_model=24, + encoder_layers=2, + decoder_layers=2, + encoder_attention_heads=2, + decoder_attention_heads=2, + encoder_ffn_dim=32, + decoder_ffn_dim=32, + max_position_embeddings=48, + eos_token_id=2, + pad_token_id=1, + bos_token_id=0, + ) + lm_model = MvpForConditionalGeneration(config).to(torch_device) + lm_model.eval() + + max_length = 5 + generated_ids = lm_model.generate( + input_ids.clone(), + do_sample=True, + num_return_sequences=1, + num_beams=2, + no_repeat_ngram_size=3, + max_length=max_length, + ) + self.assertEqual(generated_ids.shape, (input_ids.shape[0], max_length)) + + def test_shift_tokens_right(self): + input_ids = torch.tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=torch.long) + shifted = shift_tokens_right(input_ids, 1, 2) + n_pad_before = input_ids.eq(1).float().sum() + n_pad_after = shifted.eq(1).float().sum() + self.assertEqual(shifted.shape, input_ids.shape) + self.assertEqual(n_pad_after, n_pad_before - 1) + self.assertTrue(torch.eq(shifted[:, 0], 2).all()) + + @slow + def test_tokenization(self): + tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp") + examples = [" Hello world", " DomDramg"] # need leading spaces for equality + fairseq_results = [ + torch.tensor([0, 20920, 232, 2]), + torch.tensor([0, 11349, 495, 4040, 571, 2]), + ] + for ex, desired_result in zip(examples, fairseq_results): + mvp_toks = tokenizer.encode(ex, return_tensors="pt").squeeze() + assert_tensors_close(desired_result.long(), mvp_toks, prefix=ex) + + def test_generate_fp16(self): + config, input_ids, batch_size = self._get_config_and_data() + attention_mask = input_ids.ne(1).to(torch_device) + model = MvpForConditionalGeneration(config).eval().to(torch_device) + if torch_device == "cuda": + model.half() + model.generate(input_ids, attention_mask=attention_mask) + model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3) + + def test_dummy_inputs(self): + config, *_ = self._get_config_and_data() + model = MvpForConditionalGeneration(config).eval().to(torch_device) + model(**model.dummy_inputs) + + def test_resize_tokens_embeddings_more(self): + config, input_ids, _ = self._get_config_and_data() + + def _get_embs(m): + return (m.get_input_embeddings().weight.data.clone(), m.get_output_embeddings().weight.data.clone()) + + model = MvpForConditionalGeneration(config).eval().to(torch_device) + input, output = _get_embs(model) + self.assertTrue(torch.eq(input, output).all()) + new_vocab_size = 45 + model.resize_token_embeddings(new_vocab_size) + input_new, output_new = _get_embs(model) + self.assertEqual(input_new.shape, (new_vocab_size, config.d_model)) + self.assertEqual(output_new.shape, (new_vocab_size, config.d_model)) + self.assertTrue(torch.eq(input_new, output_new).all()) + + +@require_torch +class MvpModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase): + all_model_classes = ( + (MvpModel, MvpForConditionalGeneration, MvpForSequenceClassification, MvpForQuestionAnswering) + if is_torch_available() + else () + ) + all_generative_model_classes = (MvpForConditionalGeneration,) if is_torch_available() else () + is_encoder_decoder = True + fx_compatible = False + test_pruning = False + test_missing_keys = False + + def setUp(self): + self.model_tester = MvpModelTester(self) + self.config_tester = ConfigTester(self, config_class=MvpConfig) + + def test_config(self): + self.config_tester.run_common_tests() + + def test_save_load_strict(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs() + for model_class in self.all_model_classes: + model = model_class(config) + + with tempfile.TemporaryDirectory() as tmpdirname: + model.save_pretrained(tmpdirname) + model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True) + self.assertEqual(info["missing_keys"], []) + + def test_decoder_model_past_with_large_inputs(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs) + + def test_encoder_decoder_model_standalone(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common() + self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs) + + # MvpForSequenceClassification does not support inputs_embeds + def test_inputs_embeds(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + + for model_class in (MvpModel, MvpForConditionalGeneration, MvpForQuestionAnswering): + model = model_class(config) + model.to(torch_device) + model.eval() + + inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class)) + + if not self.is_encoder_decoder: + input_ids = inputs["input_ids"] + del inputs["input_ids"] + else: + encoder_input_ids = inputs["input_ids"] + decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids) + del inputs["input_ids"] + inputs.pop("decoder_input_ids", None) + + wte = model.get_input_embeddings() + if not self.is_encoder_decoder: + inputs["inputs_embeds"] = wte(input_ids) + else: + inputs["inputs_embeds"] = wte(encoder_input_ids) + inputs["decoder_inputs_embeds"] = wte(decoder_input_ids) + + with torch.no_grad(): + model(**inputs)[0] + + def test_generate_fp16(self): + config, input_dict = self.model_tester.prepare_config_and_inputs() + input_ids = input_dict["input_ids"] + attention_mask = input_ids.ne(1).to(torch_device) + model = MvpForConditionalGeneration(config).eval().to(torch_device) + if torch_device == "cuda": + model.half() + model.generate(input_ids, attention_mask=attention_mask) + model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3) + + +def assert_tensors_close(a, b, atol=1e-12, prefix=""): + """If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error.""" + if a is None and b is None: + return True + try: + if torch.allclose(a, b, atol=atol): + return True + raise + except Exception: + pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item() + if a.numel() > 100: + msg = f"tensor values are {pct_different:.1%} percent different." + else: + msg = f"{a} != {b}" + if prefix: + msg = prefix + ": " + msg + raise AssertionError(msg) + + +def _long_tensor(tok_lst): + return torch.tensor(tok_lst, dtype=torch.long, device=torch_device) + + +@require_torch +@require_sentencepiece +@require_tokenizers +class MvpModelIntegrationTests(unittest.TestCase): + @cached_property + def default_tokenizer(self): + return MvpTokenizer.from_pretrained("RUCAIBox/mvp") + + @slow + def test_inference_no_head(self): + model = MvpModel.from_pretrained("RUCAIBox/mvp").to(torch_device) + input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) + attention_mask = input_ids.ne(model.config.pad_token_id) + with torch.no_grad(): + output = model(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state + expected_shape = torch.Size((1, 11, 1024)) + self.assertEqual(output.shape, expected_shape) + expected_slice = torch.tensor( + [[0.3461, 0.3624, 0.2689], [0.3461, 0.3624, 0.2689], [-0.1562, 1.1637, -0.3784]], device=torch_device + ) + self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3)) + + @slow + def test_summarization_inference(self): + model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp").to(torch_device) + tok = self.default_tokenizer + # fmt: off + PGE_ARTICLE = """ Listen to local radio broadcasts for advertisements that reference casinos in your area.\nIf none are in your area, listen to national radio broadcasts for advertisements of casinos in other areas.\nNote the location that is mentioned in each advertisement that involves a casino.\nIf no locations are mentioned, note any additional contact information, such as a website or phone number. Use that information to find out where the casinos are.;\n,\n\nIf you learn about more than 1 casino on the radio, use the Internet to search the distance between your location and each casino. Sites such as maps.google.com or mapquest.com will help you in this search.'""" + # fmt: on + EXPECTED_SUMMARY = "Listen to the radio.\nUse the Internet." + dct = tok.batch_encode_plus( + [PGE_ARTICLE], + return_tensors="pt", + ).to(torch_device) + + hypotheses_batch = model.generate(**dct) + + decoded = tok.batch_decode(hypotheses_batch, skip_special_tokens=True) + self.assertEqual(EXPECTED_SUMMARY, decoded[0]) + + +class MvpStandaloneDecoderModelTester: + def __init__( + self, + parent, + vocab_size=99, + batch_size=13, + d_model=16, + decoder_seq_length=7, + is_training=True, + is_decoder=True, + use_attention_mask=True, + use_cache=False, + use_labels=True, + decoder_start_token_id=2, + decoder_ffn_dim=32, + decoder_layers=4, + encoder_attention_heads=4, + decoder_attention_heads=4, + max_position_embeddings=30, + is_encoder_decoder=False, + pad_token_id=0, + bos_token_id=1, + eos_token_id=2, + scope=None, + ): + self.parent = parent + self.batch_size = batch_size + self.decoder_seq_length = decoder_seq_length + # For common tests + self.seq_length = self.decoder_seq_length + self.is_training = is_training + self.use_attention_mask = use_attention_mask + self.use_labels = use_labels + + self.vocab_size = vocab_size + self.d_model = d_model + self.hidden_size = d_model + self.num_hidden_layers = decoder_layers + self.decoder_layers = decoder_layers + self.decoder_ffn_dim = decoder_ffn_dim + self.encoder_attention_heads = encoder_attention_heads + self.decoder_attention_heads = decoder_attention_heads + self.num_attention_heads = decoder_attention_heads + self.eos_token_id = eos_token_id + self.bos_token_id = bos_token_id + self.pad_token_id = pad_token_id + self.decoder_start_token_id = decoder_start_token_id + self.use_cache = use_cache + self.max_position_embeddings = max_position_embeddings + self.is_encoder_decoder = is_encoder_decoder + + self.scope = None + self.decoder_key_length = decoder_seq_length + self.base_model_out_len = 2 + self.decoder_attention_idx = 1 + + def prepare_config_and_inputs(self): + input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size) + + attention_mask = None + if self.use_attention_mask: + attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2) + + lm_labels = None + if self.use_labels: + lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size) + + config = MvpConfig( + vocab_size=self.vocab_size, + d_model=self.d_model, + encoder_layers=self.decoder_layers, + decoder_layers=self.decoder_layers, + decoder_ffn_dim=self.decoder_ffn_dim, + encoder_attention_heads=self.encoder_attention_heads, + decoder_attention_heads=self.decoder_attention_heads, + eos_token_id=self.eos_token_id, + bos_token_id=self.bos_token_id, + use_cache=self.use_cache, + pad_token_id=self.pad_token_id, + decoder_start_token_id=self.decoder_start_token_id, + max_position_embeddings=self.max_position_embeddings, + is_encoder_decoder=self.is_encoder_decoder, + ) + + return ( + config, + input_ids, + attention_mask, + lm_labels, + ) + + def prepare_config_and_inputs_for_decoder(self): + ( + config, + input_ids, + attention_mask, + lm_labels, + ) = self.prepare_config_and_inputs() + + encoder_hidden_states = floats_tensor([self.batch_size, self.decoder_seq_length, self.hidden_size]) + encoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2) + + return ( + config, + input_ids, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + lm_labels, + ) + + def create_and_check_decoder_model_past( + self, + config, + input_ids, + attention_mask, + lm_labels, + ): + config.use_cache = True + model = MvpDecoder(config=config).to(torch_device).eval() + # first forward pass + outputs = model(input_ids, use_cache=True) + outputs_use_cache_conf = model(input_ids) + outputs_no_past = model(input_ids, use_cache=False) + + self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf)) + self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1) + + past_key_values = outputs["past_key_values"] + + # create hypothetical next token and extent to next_input_ids + next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size) + + # append to next input_ids and + next_input_ids = torch.cat([input_ids, next_tokens], dim=-1) + + output_from_no_past = model(next_input_ids)["last_hidden_state"] + output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"] + + # select random slice + random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() + output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach() + output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach() + + # test that outputs are equal for slice + assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3) + + def create_and_check_decoder_model_attention_mask_past( + self, + config, + input_ids, + attention_mask, + lm_labels, + ): + model = MvpDecoder(config=config).to(torch_device).eval() + + # create attention mask + attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device) + + half_seq_length = input_ids.shape[-1] // 2 + attn_mask[:, half_seq_length:] = 0 + + # first forward pass + past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"] + + # create hypothetical next token and extent to next_input_ids + next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size) + + # change a random masked slice from input_ids + random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1 + random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1) + input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens + + # append to next input_ids and attn_mask + next_input_ids = torch.cat([input_ids, next_tokens], dim=-1) + attn_mask = torch.cat( + [attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)], + dim=1, + ) + + # get two different outputs + output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"] + output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[ + "last_hidden_state" + ] + + # select random slice + random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() + output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach() + output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach() + + # test that outputs are equal for slice + assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3) + + def prepare_config_and_inputs_for_common(self): + config_and_inputs = self.prepare_config_and_inputs() + ( + config, + input_ids, + attention_mask, + lm_labels, + ) = config_and_inputs + + inputs_dict = { + "input_ids": input_ids, + "attention_mask": attention_mask, + } + return config, inputs_dict + + +@require_torch +class MvpStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase): + all_model_classes = (MvpDecoder, MvpForCausalLM) if is_torch_available() else () + all_generative_model_classes = (MvpForCausalLM,) if is_torch_available() else () + fx_comptatible = True + test_pruning = False + is_encoder_decoder = False + + def setUp( + self, + ): + self.model_tester = MvpStandaloneDecoderModelTester(self, is_training=False) + self.config_tester = ConfigTester(self, config_class=MvpConfig) + + def test_config(self): + self.config_tester.run_common_tests() + + def test_decoder_model_past(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_decoder_model_past(*config_and_inputs) + + def test_decoder_model_attn_mask_past(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs) + + def test_retain_grad_hidden_states_attentions(self): + # decoder cannot keep gradients + return diff --git a/tests/models/mvp/test_tokenization_mvp.py b/tests/models/mvp/test_tokenization_mvp.py new file mode 100644 index 0000000000..ad3fad67c9 --- /dev/null +++ b/tests/models/mvp/test_tokenization_mvp.py @@ -0,0 +1,185 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# 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. +import json +import os +import unittest + +from transformers import BatchEncoding, MvpTokenizer, MvpTokenizerFast +from transformers.models.roberta.tokenization_roberta import VOCAB_FILES_NAMES +from transformers.testing_utils import require_tokenizers, require_torch +from transformers.utils import cached_property + +from ...test_tokenization_common import TokenizerTesterMixin, filter_roberta_detectors + + +@require_tokenizers +class TestTokenizationMvp(TokenizerTesterMixin, unittest.TestCase): + tokenizer_class = MvpTokenizer + rust_tokenizer_class = MvpTokenizerFast + test_rust_tokenizer = True + from_pretrained_filter = filter_roberta_detectors + # from_pretrained_kwargs = {'add_prefix_space': True} + + def setUp(self): + super().setUp() + vocab = [ + "l", + "o", + "w", + "e", + "r", + "s", + "t", + "i", + "d", + "n", + "\u0120", + "\u0120l", + "\u0120n", + "\u0120lo", + "\u0120low", + "er", + "\u0120lowest", + "\u0120newer", + "\u0120wider", + "", + ] + vocab_tokens = dict(zip(vocab, range(len(vocab)))) + merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""] + self.special_tokens_map = {"unk_token": ""} + + self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"]) + self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"]) + with open(self.vocab_file, "w", encoding="utf-8") as fp: + fp.write(json.dumps(vocab_tokens) + "\n") + with open(self.merges_file, "w", encoding="utf-8") as fp: + fp.write("\n".join(merges)) + + def get_tokenizer(self, **kwargs): + kwargs.update(self.special_tokens_map) + return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs) + + def get_rust_tokenizer(self, **kwargs): + kwargs.update(self.special_tokens_map) + return self.rust_tokenizer_class.from_pretrained(self.tmpdirname, **kwargs) + + def get_input_output_texts(self, tokenizer): + return "lower newer", "lower newer" + + @cached_property + def default_tokenizer(self): + return MvpTokenizer.from_pretrained("RUCAIBox/mvp") + + @cached_property + def default_tokenizer_fast(self): + return MvpTokenizerFast.from_pretrained("RUCAIBox/mvp") + + @require_torch + def test_prepare_batch(self): + src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."] + expected_src_tokens = [0, 250, 251, 17818, 13, 39186, 1938, 4, 2] + + for tokenizer in [self.default_tokenizer, self.default_tokenizer_fast]: + batch = tokenizer(src_text, max_length=len(expected_src_tokens), padding=True, return_tensors="pt") + self.assertIsInstance(batch, BatchEncoding) + + self.assertEqual((2, 9), batch.input_ids.shape) + self.assertEqual((2, 9), batch.attention_mask.shape) + result = batch.input_ids.tolist()[0] + self.assertListEqual(expected_src_tokens, result) + # Test that special tokens are reset + + @require_torch + def test_prepare_batch_empty_target_text(self): + src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."] + for tokenizer in [self.default_tokenizer, self.default_tokenizer_fast]: + batch = tokenizer(src_text, padding=True, return_tensors="pt") + # check if input_ids are returned and no labels + self.assertIn("input_ids", batch) + self.assertIn("attention_mask", batch) + self.assertNotIn("labels", batch) + self.assertNotIn("decoder_attention_mask", batch) + + @require_torch + def test_as_target_tokenizer_target_length(self): + tgt_text = [ + "Summary of the text.", + "Another summary.", + ] + for tokenizer in [self.default_tokenizer, self.default_tokenizer_fast]: + with tokenizer.as_target_tokenizer(): + targets = tokenizer(tgt_text, max_length=32, padding="max_length", return_tensors="pt") + self.assertEqual(32, targets["input_ids"].shape[1]) + + @require_torch + def test_prepare_batch_not_longer_than_maxlen(self): + for tokenizer in [self.default_tokenizer, self.default_tokenizer_fast]: + batch = tokenizer( + ["I am a small frog" * 1024, "I am a small frog"], padding=True, truncation=True, return_tensors="pt" + ) + self.assertIsInstance(batch, BatchEncoding) + self.assertEqual(batch.input_ids.shape, (2, 1024)) + + @require_torch + def test_special_tokens(self): + + src_text = ["A long paragraph for summarization."] + tgt_text = [ + "Summary of the text.", + ] + for tokenizer in [self.default_tokenizer, self.default_tokenizer_fast]: + inputs = tokenizer(src_text, return_tensors="pt") + with tokenizer.as_target_tokenizer(): + targets = tokenizer(tgt_text, return_tensors="pt") + input_ids = inputs["input_ids"] + labels = targets["input_ids"] + self.assertTrue((input_ids[:, 0] == tokenizer.bos_token_id).all().item()) + self.assertTrue((labels[:, 0] == tokenizer.bos_token_id).all().item()) + self.assertTrue((input_ids[:, -1] == tokenizer.eos_token_id).all().item()) + self.assertTrue((labels[:, -1] == tokenizer.eos_token_id).all().item()) + + def test_pretokenized_inputs(self): + pass + + def test_embeded_special_tokens(self): + for tokenizer, pretrained_name, kwargs in self.tokenizers_list: + with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"): + tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs) + tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs) + sentence = "A, AllenNLP sentence." + tokens_r = tokenizer_r.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True) + tokens_p = tokenizer_p.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True) + + # token_type_ids should put 0 everywhere + self.assertEqual(sum(tokens_r["token_type_ids"]), sum(tokens_p["token_type_ids"])) + + # attention_mask should put 1 everywhere, so sum over length should be 1 + self.assertEqual( + sum(tokens_r["attention_mask"]) / len(tokens_r["attention_mask"]), + sum(tokens_p["attention_mask"]) / len(tokens_p["attention_mask"]), + ) + + tokens_r_str = tokenizer_r.convert_ids_to_tokens(tokens_r["input_ids"]) + tokens_p_str = tokenizer_p.convert_ids_to_tokens(tokens_p["input_ids"]) + + # Rust correctly handles the space before the mask while python doesnt + self.assertSequenceEqual(tokens_p["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2]) + self.assertSequenceEqual(tokens_r["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2]) + + self.assertSequenceEqual( + tokens_p_str, ["", "A", ",", "", "ĠAllen", "N", "LP", "Ġsentence", ".", ""] + ) + self.assertSequenceEqual( + tokens_r_str, ["", "A", ",", "", "ĠAllen", "N", "LP", "Ġsentence", ".", ""] + ) diff --git a/utils/check_repo.py b/utils/check_repo.py index 20d6b136d1..9905bb0054 100644 --- a/utils/check_repo.py +++ b/utils/check_repo.py @@ -78,6 +78,8 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [ "MegatronBertEncoder", # Building part of bigger (tested) model. "MegatronBertDecoder", # Building part of bigger (tested) model. "MegatronBertDecoderWrapper", # Building part of bigger (tested) model. + "MvpDecoderWrapper", # Building part of bigger (tested) model. + "MvpEncoder", # Building part of bigger (tested) model. "PegasusEncoder", # Building part of bigger (tested) model. "PegasusDecoderWrapper", # Building part of bigger (tested) model. "DPREncoder", # Building part of bigger (tested) model.