[Time-Series] informer model (#21099)

* added informer to gitignore

* added informer to gitignore

* WIP informer2020

* added checking that instantiate works

* added config using gluonTS by kashif

* WIP config

* adding informeConfig. need to remove FeatureEmbedder

* done InformerConfig, but need to change the names

* Done informer model init. working on enc-dec

* added things to address, after reading again enc-dec in the paper

* done modeling - checking initialization work

* added informer to gitignore

* WIP informer2020

* added checking that instantiate works

* added config using gluonTS by kashif

* WIP config

* adding informeConfig. need to remove FeatureEmbedder

* done InformerConfig, but need to change the names

* Done informer model init. working on enc-dec

* added things to address, after reading again enc-dec in the paper

* done modeling - checking initialization work

* moved enc-dec init to InformerEncoder/Decoder init

* added 'init_std' to config, now model init works!

* WIP conversion script, and added code sources

* WIP conversion script: loading original informer pth works

* WIP conversion script: change defaults in the config

* WIP conversion script: supporting Informer input embedding

* WIP conversion script: added parameters for the informer embed

* WIP conversion script: change dim_feedforward=2048

* WIP conversion script: remove unused args for loading checkpoint

* just cleaning up

* DataEmbedding removed, after thinking with Kashif

* working on forward pass

* WIP forward pass: trying to establish working batch for forward pass

* cleaning and finalizing

* adding HF names and docs

* init after cleaning works

* WIP in tests

* added docs for the informer specific args

* fix style

* undo change

* cleaning informer, now need to work only enc-dec

* initial enc-dec classes

* added encoder and decoder

* added todo

* add todos for conv_layers

* added decoder docs from vanilla

* added encoder docs from vanilla

* remove encoder decoder from the original informer

* removed AttentionLayer from the original paper

* removed TriangularCausalMask, same as decoder_attention_mask

* initial sparse attention

* use conv_layers

* fixed test_config test

* fix parenthesis when itearting zip(layers, conv_layers)

* error found in prob attention, added sizes as comments

* fix sizes

* added proposal for q_reduce indexing, and remove unused

* WIP ProbMask, and changed factor=2 for testing

* remove unused libs for this PR for creating the env

* fix checking the attn_weights.size() after bmm

* Q_reduce: changed from torch.gather to simple slicing

* WIP calculate final attn_output

* finish adding v_aggregated, attn_output ready

* changed tgt_len to u in attention_mask, need to fix the size error

* comment attention_mask for encoder, and fix if cond for v_agg

* added ProbMask support (wip), removed old original code

* finished ProbMask 😃



* Revert "remove unused libs for this PR for creating the env"

This reverts commit 11a081e09e92771e51a5d2758d53a9afb59547f0.

* fixes

* make style

* fix initial tests

* fix more tests

* dry

* make style

* remove unused files

* style

* added integration tests

* fix num_static_real_features

* fix header

* remove unused function

* fix example

* fix docs

* Update src/transformers/models/informer/configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/informer/modeling_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/informer/configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/informer/configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/informer/configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/informer/configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixes for reviewer

* use prediction_length from model

* fix style

* fixed informer.mdx

* added to index

* updated readme

* undo

* make fix-copies

* typo

* fix copy

* added Informer to toctree

* in order

* fixed comments

* remove unneeded new lines in docs

* make static real and cat optional

* fix use of distil conv layers

* fixed integration test

* added checkpoint for convlayer

* make fix-copies

* updated from time series model

* make fix-copies

* copy decoder

* fix unit tests

* updated scaling config

* fix integration tests

* IGNORE_NON_TESTED

* IGNORE_NON_AUTO_CONFIGURED

* IGNORE_NON_AUTO_CONFIGURED

* updated check configs

* fix formatting

* undo change from time series

* prediction_length should not be None

* aliign with the blog: prettify ProbSparse and change attention_factor  to sampling_factor

* make style

* make fix-copies

* niels CR: update contributed by

* niels CR: update configuration_informer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* niels CR: update kashif -> huggingface
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* niels CR: `sampling_factor` only relevant when `attention_type`=prob

* make style

* fixed U_part: added multiplication by `L_Q`

* fixed bug: remove `is not None` from `if config.distil`

* fixed test: `decoder_seq_length` to `encoder_seq_length` in cross_attentions check

* fix integration tests

* updated model hub

* do not shift as in training

* undo

* fix make-copies

* make fix-copies

* added `if prediction_length is None`

* changed `ProbSparseAttention` to `InformerProbSparseAttention`

* changed `V_sum` -> `v_mean_dim_time`

* changed `ConvLayer` to `InformerConvLayer` and fixed `super()`

* TimeSeriesTansformer->Informer in decoder's Copied from

* more descriptive in ProbSparse

* make style

* fix coped from

* Revert "added `if prediction_length is None`"

This reverts commit b4cbddfa05e3bd739b79569cd3c3b89e316f2451.

* fixed indent

* use InformerSinusoidalPositionalEmbedding

* make fix-style

* fix from #21860

* fix name

* make fix-copies

* use time series utils

* fix dec num_heads

* docstring

* added time series util doc

* _import_structure

* formatting

* changes from review

* make style

* fix docs

* fix doc

* removed NegativeLogLikelihood

---------
Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

README.md

@@ -354,6 +354,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.

README_es.md

@@ -342,6 +342,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.

README_hd.md

@@ -314,6 +314,7 @@ conda install -c huggingface transformers
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (फेसबुक से) साथ में पेपर [ह्यूबर्ट: सेल्फ सुपरवाइज्ड स्पीच रिप्रेजेंटेशन लर्निंग बाय मास्क्ड प्रेडिक्शन ऑफ हिडन यूनिट्स](https ://arxiv.org/abs/2106.07447) वेई-निंग सू, बेंजामिन बोल्टे, याओ-हंग ह्यूबर्ट त्साई, कुशाल लखोटिया, रुस्लान सालाखुतदीनोव, अब्देलरहमान मोहम्मद द्वारा।
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (बर्कले से) साथ में कागज [I-BERT: Integer-only BERT Quantization](https:// arxiv.org/abs/2101.01321) सेहून किम, अमीर घोलमी, ज़ेवेई याओ, माइकल डब्ल्यू महोनी, कर्ट केटज़र द्वारा।
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.

README_ja.md

@@ -376,6 +376,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook から) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed から公開された研究論文: [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447)
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley から) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer から公開された研究論文: [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321)
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI から) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever から公開された研究論文: [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/)
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI から) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever から公開された研究論文: [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf)
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia から) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou から公開された研究論文: [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318)
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (Microsoft Research Asia から) Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou から公開された研究論文: [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740)

README_ko.md

@@ -291,6 +291,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook 에서) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 의 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 논문과 함께 발표했습니다.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley 에서) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 의 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 논문과 함께 발표했습니다.
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI 에서) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 의 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 논문과 함께 발표했습니다.
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI 에서) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever 의 [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) 논문과 함께 발표했습니다.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia 에서) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 의 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 논문과 함께 발표했습니다.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (Microsoft Research Asia 에서) Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou 의 [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) 논문과 함께 발표했습니다.

README_zh-hans.md

@@ -315,6 +315,7 @@ conda install -c huggingface transformers
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (来自 Facebook) 伴随论文 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 由 Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 发布。
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (来自 Berkeley) 伴随论文 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 由 Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 发布。
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (来自 OpenAI) 伴随论文 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 由 Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 发布。
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 由 Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 发布。
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) 由 Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou 发布。

README_zh-hant.md

@@ -327,6 +327,7 @@ conda install -c huggingface transformers
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](https://huggingface.co/docs/transformers/main/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.

docs/source/en/_toctree.yml

@@ -614,6 +614,8 @@
       title: Reinforcement learning models
     - isExpanded: false
       sections:
+      - local: model_doc/informer
+        title: Informer
       - local: model_doc/time_series_transformer
         title: Time Series Transformer
       title: Time series models
@@ -640,5 +642,7 @@
       title: Utilities for Audio processing
     - local: internal/file_utils
       title: General Utilities
+    - local: internal/time_series_utils
+      title: Utilities for Time Series
     title: Internal Helpers
   title: API

docs/source/en/index.mdx

@@ -128,6 +128,7 @@ The documentation is organized into five sections:
 1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[ImageGPT](model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[Jukebox](model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
 1. **[LayoutLM](model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
@@ -318,6 +319,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            Hubert             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            I-BERT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           ImageGPT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           Informer            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            Jukebox            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           LayoutLM            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |          LayoutLMv2           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |

docs/source/en/internal/time_series_utils.mdx

+<!--Copyright 2023 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.
+-->
+
+# Time Series Utilities
+
+This page lists all the utility functions and classes that can be used for Time Series based models.
+
+Most of those are only useful if you are studying the code of the time series models or you wish to add to the collection of distributional output classes.
+
+## Distributional Output
+
+[[autodoc]] time_series_utils.NormalOutput
+
+[[autodoc]] time_series_utils.StudentTOutput
+
+[[autodoc]] time_series_utils.NegativeBinomialOutput

docs/source/en/main_classes/output.mdx

@@ -164,6 +164,18 @@ documented on their corresponding model page.
 
 [[autodoc]] modeling_outputs.XVectorOutput
 
+## Seq2SeqTSModelOutput
+
+[[autodoc]] modeling_outputs.Seq2SeqTSModelOutput
+
+## Seq2SeqTSPredictionOutput
+
+[[autodoc]] modeling_outputs.Seq2SeqTSPredictionOutput
+
+## SampleTSPredictionOutput
+
+[[autodoc]] modeling_outputs.SampleTSPredictionOutput
+
 ## TFBaseModelOutput
 
 [[autodoc]] modeling_tf_outputs.TFBaseModelOutput

docs/source/en/model_doc/informer.mdx

+<!--Copyright 2023 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.
+-->
+
+# Informer
+
+## Overview
+
+The Informer model was proposed in [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting ](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
+
+This method introduces a Probabilistic Attention mechanism to select the "active" queries rather than the "lazy" queries and provides a sparse Transformer thus mitigating the quadratic compute and memory requirements of vanilla attention.
+
+The abstract from the paper is the following:
+
+*Many real-world applications require the prediction of long sequence time-series, such as electricity consumption planning. Long sequence time-series forecasting (LSTF) demands a high prediction capacity of the model, which is the ability to capture precise long-range dependency coupling between output and input efficiently. Recent studies have shown the potential of Transformer to increase the prediction capacity. However, there are several severe issues with Transformer that prevent it from being directly applicable to LSTF, including quadratic time complexity, high memory usage, and inherent limitation of the encoder-decoder architecture. To address these issues, we design an efficient transformer-based model for LSTF, named Informer, with three distinctive characteristics: (i) a ProbSparse self-attention mechanism, which achieves O(L logL) in time complexity and memory usage, and has comparable performance on sequences' dependency alignment. (ii) the self-attention distilling highlights dominating attention by halving cascading layer input, and efficiently handles extreme long input sequences. (iii) the generative style decoder, while conceptually simple, predicts the long time-series sequences at one forward operation rather than a step-by-step way, which drastically improves the inference speed of long-sequence predictions. Extensive experiments on four large-scale datasets demonstrate that Informer significantly outperforms existing methods and provides a new solution to the LSTF problem.*
+
+This model was contributed by [elisim](https://huggingface.co/elisim) and [kashif](https://huggingface.co/kashif).
+The original code can be found [here](https://github.com/zhouhaoyi/Informer2020).
+
+
+## InformerConfig
+
+[[autodoc]] InformerConfig
+
+
+## InformerModel
+
+[[autodoc]] InformerModel
+    - forward
+
+
+## InformerForPrediction
+
+[[autodoc]] InformerForPrediction
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -318,6 +318,7 @@ _import_structure = {
     "models.hubert": ["HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "HubertConfig"],
     "models.ibert": ["IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "IBertConfig"],
     "models.imagegpt": ["IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ImageGPTConfig"],
+    "models.informer": ["INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "InformerConfig"],
     "models.jukebox": [
         "JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "JukeboxConfig",
@@ -1706,6 +1707,14 @@ else:
             "load_tf_weights_in_imagegpt",
         ]
     )
+    _import_structure["models.informer"].extend(
+        [
+            "INFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "InformerForPrediction",
+            "InformerModel",
+            "InformerPreTrainedModel",
+        ]
+    )
     _import_structure["models.jukebox"].extend(
         [
             "JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2698,6 +2707,7 @@ else:
     ]
     _import_structure["pytorch_utils"] = ["Conv1D", "apply_chunking_to_forward", "prune_layer"]
     _import_structure["sagemaker"] = []
+    _import_structure["time_series_utils"] = []
     _import_structure["trainer"] = ["Trainer"]
     _import_structure["trainer_pt_utils"] = ["torch_distributed_zero_first"]
     _import_structure["trainer_seq2seq"] = ["Seq2SeqTrainer"]
@@ -3900,6 +3910,7 @@ if TYPE_CHECKING:
     from .models.hubert import HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, HubertConfig
     from .models.ibert import IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, IBertConfig
     from .models.imagegpt import IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP, ImageGPTConfig
+    from .models.informer import INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, InformerConfig
     from .models.jukebox import (
         JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP,
         JukeboxConfig,
@@ -5079,6 +5090,12 @@ if TYPE_CHECKING:
             ImageGPTPreTrainedModel,
             load_tf_weights_in_imagegpt,
         )
+        from .models.informer import (
+            INFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            InformerForPrediction,
+            InformerModel,
+            InformerPreTrainedModel,
+        )
         from .models.jukebox import (
             JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST,
             JukeboxModel,

src/transformers/modeling_outputs.py

@@ -1464,3 +1464,161 @@ class Seq2SeqSpectrogramOutput(ModelOutput):
     encoder_last_hidden_state: Optional[torch.FloatTensor] = None
     encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
     encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class Seq2SeqTSModelOutput(ModelOutput):
+    """
+    Base class for time series model's encoder outputs that also contains pre-computed hidden states that can speed up
+    sequential decoding.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+
+            If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
+            hidden_size)` is output.
+        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.
+        decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the optional initial embedding outputs.
+        decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (`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 of the model.
+        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the optional initial embedding outputs.
+        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        loc (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
+            Shift values of each time series' context window which is used to give the model inputs of the same
+            magnitude and then used to shift back to the original magnitude.
+        scale (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
+            Scaling values of each time series' context window which is used to give the model inputs of the same
+            magnitude and then used to rescale back to the original magnitude.
+        static_features: (`torch.FloatTensor` of shape `(batch_size, feature size)`, *optional*):
+            Static features of each time series' in a batch which are copied to the covariates at inference time.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    loc: Optional[torch.FloatTensor] = None
+    scale: Optional[torch.FloatTensor] = None
+    static_features: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class Seq2SeqTSPredictionOutput(ModelOutput):
+    """
+    Base class for time series model's decoder outputs that also contain the loss as well as the parameters of the
+    chosen distribution.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when a `future_values` is provided):
+            Distributional loss.
+        params (`torch.FloatTensor` of shape `(batch_size, num_samples, num_params)`):
+            Parameters of the chosen distribution.
+        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.
+        decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (`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 of the model.
+        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        loc (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
+            Shift values of each time series' context window which is used to give the model inputs of the same
+            magnitude and then used to shift back to the original magnitude.
+        scale (`torch.FloatTensor` of shape `(batch_size,)` or `(batch_size, input_size)`, *optional*):
+            Scaling values of each time series' context window which is used to give the model inputs of the same
+            magnitude and then used to rescale back to the original magnitude.
+        static_features: (`torch.FloatTensor` of shape `(batch_size, feature size)`, *optional*):
+            Static features of each time series' in a batch which are copied to the covariates at inference time.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    params: Optional[Tuple[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    loc: Optional[torch.FloatTensor] = None
+    scale: Optional[torch.FloatTensor] = None
+    static_features: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class SampleTSPredictionOutput(ModelOutput):
+    """
+    Base class for time series model's predictions outputs that contains the sampled values from the chosen
+    distribution.
+
+    Args:
+        sequences (`torch.FloatTensor` of shape `(batch_size, num_samples, prediction_length)` or `(batch_size, num_samples, prediction_length, input_size)`):
+            Sampled values from the chosen distribution.
+    """
+
+    sequences: torch.FloatTensor = None

src/transformers/models/__init__.py

@@ -93,6 +93,7 @@ from . import (
     hubert,
     ibert,
     imagegpt,
+    informer,
     jukebox,
     layoutlm,
     layoutlmv2,

src/transformers/models/auto/configuration_auto.py

@@ -100,6 +100,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("hubert", "HubertConfig"),
         ("ibert", "IBertConfig"),
         ("imagegpt", "ImageGPTConfig"),
+        ("informer", "InformerConfig"),
         ("jukebox", "JukeboxConfig"),
         ("layoutlm", "LayoutLMConfig"),
         ("layoutlmv2", "LayoutLMv2Config"),
@@ -274,6 +275,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("hubert", "HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("ibert", "IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("imagegpt", "IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("informer", "INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("jukebox", "JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("layoutlm", "LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("layoutlmv2", "LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -450,6 +452,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("hubert", "Hubert"),
         ("ibert", "I-BERT"),
         ("imagegpt", "ImageGPT"),
+        ("informer", "Informer"),
         ("jukebox", "Jukebox"),
         ("layoutlm", "LayoutLM"),
         ("layoutlmv2", "LayoutLMv2"),

src/transformers/models/auto/modeling_auto.py

@@ -98,6 +98,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("hubert", "HubertModel"),
         ("ibert", "IBertModel"),
         ("imagegpt", "ImageGPTModel"),
+        ("informer", "InformerModel"),
         ("jukebox", "JukeboxModel"),
         ("layoutlm", "LayoutLMModel"),
         ("layoutlmv2", "LayoutLMv2Model"),

src/transformers/models/informer/__init__.py

+# Copyright 2023 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
+
+# rely on isort to merge the imports
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_informer": [
+        "INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "InformerConfig",
+    ],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_informer"] = [
+        "INFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "InformerForPrediction",
+        "InformerModel",
+        "InformerPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_informer import INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, InformerConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_informer import (
+            INFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            InformerForPrediction,
+            InformerModel,
+            InformerPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/informer/configuration_informer.py

+# coding=utf-8
+# Copyright 2023 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.
+"""Informer model configuration"""
+
+from typing import List, Optional, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "huggingface/informer-tourism-monthly": (
+        "https://huggingface.co/huggingface/informer-tourism-monthly/resolve/main/config.json"
+    ),
+    # See all Informer models at https://huggingface.co/models?filter=informer
+}
+
+
+class InformerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of an [`InformerModel`]. It is used to instantiate an
+    Informer 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 Informer
+    [huggingface/informer-tourism-monthly](https://huggingface.co/huggingface/informer-tourism-monthly) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        prediction_length (`int`):
+            The prediction length for the decoder. In other words, the prediction horizon of the model. This value is
+            typically dictated by the dataset and we recommend to set it appropriately.
+        context_length (`int`, *optional*, defaults to `prediction_length`):
+            The context length for the encoder. If `None`, the context length will be the same as the
+            `prediction_length`.
+        distribution_output (`string`, *optional*, defaults to `"student_t"`):
+            The distribution emission head for the model. Could be either "student_t", "normal" or "negative_binomial".
+        loss (`string`, *optional*, defaults to `"nll"`):
+            The loss function for the model corresponding to the `distribution_output` head. For parametric
+            distributions it is the negative log likelihood (nll) - which currently is the only supported one.
+        input_size (`int`, *optional*, defaults to 1):
+            The size of the target variable which by default is 1 for univariate targets. Would be > 1 in case of
+            multivariate targets.
+        scaling (`string` or `bool`, *optional* defaults to `"mean"`):
+            Whether to scale the input targets via "mean" scaler, "std" scaler or no scaler if `None`. If `True`, the
+            scaler is set to "mean".
+        lags_sequence (`list[int]`, *optional*, defaults to `[1, 2, 3, 4, 5, 6, 7]`):
+            The lags of the input time series as covariates often dictated by the frequency of the data. Default is
+            `[1, 2, 3, 4, 5, 6, 7]` but we recommend to change it based on the dataset appropriately.
+        num_time_features (`int`, *optional*, defaults to 0):
+            The number of time features in the input time series.
+        num_dynamic_real_features (`int`, *optional*, defaults to 0):
+            The number of dynamic real valued features.
+        num_static_categorical_features (`int`, *optional*, defaults to 0):
+            The number of static categorical features.
+        num_static_real_features (`int`, *optional*, defaults to 0):
+            The number of static real valued features.
+        cardinality (`list[int]`, *optional*):
+            The cardinality (number of different values) for each of the static categorical features. Should be a list
+            of integers, having the same length as `num_static_categorical_features`. Cannot be `None` if
+            `num_static_categorical_features` is > 0.
+        embedding_dimension (`list[int]`, *optional*):
+            The dimension of the embedding for each of the static categorical features. Should be a list of integers,
+            having the same length as `num_static_categorical_features`. Cannot be `None` if
+            `num_static_categorical_features` is > 0.
+        d_model (`int`, *optional*, defaults to 64):
+            Dimensionality of the transformer layers.
+        encoder_layers (`int`, *optional*, defaults to 2):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 2):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 2):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 2):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 32):
+            Dimension of the "intermediate" (often named feed-forward) layer in encoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 32):
+            Dimension 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 decoder. If string, `"gelu"` and
+            `"relu"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the encoder, and decoder.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.1):
+            The dropout probability for the attention and fully connected layers for each encoder layer.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.1):
+            The dropout probability for the attention and fully connected layers for each decoder layer.
+        attention_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability used between the two layers of the feed-forward networks.
+        num_parallel_samples (`int`, *optional*, defaults to 100):
+            The number of samples to generate in parallel for each time step of inference.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated normal weight initialization distribution.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether to use the past key/values attentions (if applicable to the model) to speed up decoding.
+        attention_type (`str`, *optional*, defaults to "prob"):
+            Attention used in encoder. This can be set to "prob" (Informer's ProbAttention) or "full" (vanilla
+            transformer's canonical self-attention).
+        sampling_factor (`int`, *optional*, defaults to 5):
+            ProbSparse sampling factor (only makes affect when `attention_type`="prob"). It is used to control the
+            reduced query matrix (Q_reduce) input length.
+        distil (`bool`, *optional*, defaults to `True`):
+            Whether to use distilling in encoder.
+
+    Example:
+
+    ```python
+    >>> from transformers import InformerConfig, InformerModel
+
+    >>> # Initializing an Informer configuration with 12 time steps for prediction
+    >>> configuration = InformerConfig(prediction_length=12)
+
+    >>> # Randomly initializing a model (with random weights) from the configuration
+    >>> model = InformerModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "informer"
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+        "num_hidden_layers": "encoder_layers",
+    }
+
+    def __init__(
+        self,
+        prediction_length: Optional[int] = None,
+        context_length: Optional[int] = None,
+        distribution_output: str = "student_t",
+        loss: str = "nll",
+        input_size: int = 1,
+        lags_sequence: List[int] = None,
+        scaling: Optional[Union[str, bool]] = "mean",
+        num_dynamic_real_features: int = 0,
+        num_static_real_features: int = 0,
+        num_static_categorical_features: int = 0,
+        num_time_features: int = 0,
+        cardinality: Optional[List[int]] = None,
+        embedding_dimension: Optional[List[int]] = None,
+        d_model: int = 64,
+        encoder_ffn_dim: int = 32,
+        decoder_ffn_dim: int = 32,
+        encoder_attention_heads: int = 2,
+        decoder_attention_heads: int = 2,
+        encoder_layers: int = 2,
+        decoder_layers: int = 2,
+        is_encoder_decoder: bool = True,
+        activation_function: str = "gelu",
+        dropout: float = 0.05,
+        encoder_layerdrop: float = 0.1,
+        decoder_layerdrop: float = 0.1,
+        attention_dropout: float = 0.1,
+        activation_dropout: float = 0.1,
+        num_parallel_samples: int = 100,
+        init_std: float = 0.02,
+        use_cache=True,
+        # Informer arguments
+        attention_type: str = "prob",
+        sampling_factor: int = 5,
+        distil: bool = True,
+        **kwargs,
+    ):
+        # time series specific configuration
+        self.prediction_length = prediction_length
+        self.context_length = context_length or prediction_length
+        self.distribution_output = distribution_output
+        self.loss = loss
+        self.input_size = input_size
+        self.num_time_features = num_time_features
+        self.lags_sequence = lags_sequence if lags_sequence is not None else [1, 2, 3, 4, 5, 6, 7]
+        self.scaling = scaling
+        self.num_dynamic_real_features = num_dynamic_real_features
+        self.num_static_real_features = num_static_real_features
+        self.num_static_categorical_features = num_static_categorical_features
+
+        # set cardinality
+        if cardinality and num_static_categorical_features > 0:
+            if len(cardinality) != num_static_categorical_features:
+                raise ValueError(
+                    "The cardinality should be a list of the same length as `num_static_categorical_features`"
+                )
+            self.cardinality = cardinality
+        else:
+            self.cardinality = [0]
+
+        # set embedding_dimension
+        if embedding_dimension and num_static_categorical_features > 0:
+            if len(embedding_dimension) != num_static_categorical_features:
+                raise ValueError(
+                    "The embedding dimension should be a list of the same length as `num_static_categorical_features`"
+                )
+            self.embedding_dimension = embedding_dimension
+        else:
+            self.embedding_dimension = [min(50, (cat + 1) // 2) for cat in self.cardinality]
+
+        self.num_parallel_samples = num_parallel_samples
+
+        # Transformer architecture configuration
+        self.feature_size = input_size * len(self.lags_sequence) + self._number_of_features
+        self.d_model = d_model
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.decoder_layers = decoder_layers
+
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+
+        self.activation_function = activation_function
+        self.init_std = init_std
+
+        self.output_attentions = False
+        self.output_hidden_states = False
+
+        self.use_cache = use_cache
+
+        # Informer
+        self.attention_type = attention_type
+        self.sampling_factor = sampling_factor
+        self.distil = distil
+
+        super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
+
+    @property
+    def _number_of_features(self) -> int:
+        return (
+            sum(self.embedding_dimension)
+            + self.num_dynamic_real_features
+            + self.num_time_features
+            + self.num_static_real_features
+            + self.input_size * 2  # the log1p(abs(loc)) and log(scale) features
+        )