time series forecasting model (#17965)

* initial files * initial model via cli * typos * make a start on the model config * ready with configuation * remove tokenizer ref. * init the transformer * added initial model forward to return dec_output * require gluonts * update dep. ver table and add as extra * fixed typo * add type for prediction_length * use num_time_features * use config * more config * typos * opps another typo * freq can be none * default via transformation is 1 * initial transformations * fix imports * added transform_start_field * add helper to create pytorch dataloader * added inital val and test data loader * added initial distr head and loss * training working * remove TimeSeriesTransformerTokenizer Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * Update src/transformers/__init__.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/__init__.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * fixed copyright * removed docs * remove time series tokenizer * fixed docs * fix text * fix second * fix default * fix order * use config directly * undo change * fix comment * fix year * fix import * add additional arguments for training vs. test * initial greedy inference loop * fix inference * comment out token inputs to enc dec * Use HF encoder/decoder * fix inference * Use Seq2SeqTSModelOutput output * return Seq2SeqTSPredictionOutput * added default arguments * fix return_dict true * scale is a tensor * output static_features for inference * clean up some unused bits * fixed typo * set return_dict if none * call model once for both train/predict * use cache if future_target is none * initial generate func * generate arguments * future_time_feat is required * return SampleTSPredictionOutput * removed unneeded classes * fix when params is none * fix return dict * fix num_attention_heads * fix arguments * remove unused shift_tokens_right * add different dropout configs * implement FeatureEmbedder, Scaler and weighted_average * remove gluonts dependency * fix class names * avoid _variable names * remove gluonts dependency * fix imports * remove gluonts from configuration * fix docs * fixed typo * move utils to examples * add example requirements * config has no freq * initial run_ts_no_trainer * remove from ignore * fix output_attentions and removed unsued getters/setters * removed unsed tests * add dec seq len * add test_attention_outputs * set has_text_modality=False * add config attribute_map * make style * make fix-copies * add encoder_outputs to TimeSeriesTransformerForPrediction forward * Improve docs, add model to README * added test_forward_signature * More improvements * Add more copied from * Fix README * Fix remaining quality issues * updated encoder and decoder * fix generate * output_hidden_states and use_cache are optional * past key_values returned too * initialize weights of distribution_output module * fixed more tests * update test_forward_signature * fix return_dict outputs * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * removed commented out tests * added neg. bin and normal output * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * move to one line * Add docstrings * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * add try except for assert and raise * try and raise exception * fix the documentation formatting * fix assert call * fix docstring formatting * removed input_ids from DOCSTRING * Update input docstring * Improve variable names * Update order of inputs * Improve configuration * Improve variable names * Improve docs * Remove key_length from tests * Add extra docs * initial unittests * added test_inference_no_head test * added test_inference_head * add test_seq_to_seq_generation * make style * one line * assert mean prediction * removed comments * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> * fix order of args * make past_observed_mask optional as well * added Amazon license header * updated utils with new fieldnames * make style * cleanup * undo position of past_observed_mask * fix import * typo * more typo * rename example files * remove example for now * Update docs/source/en/_toctree.yml Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/configuration_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/time_series_transformer/modeling_time_series_transformer.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update modeling_time_series_transformer.py fix style * fixed typo * fix typo and grammer * fix style Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com> Co-authored-by: NielsRogge <niels.rogge1@gmail.com> Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
2022-09-30 21:32:59 +02:00
parent cfb777f27c
commit 5cd16f01db
19 changed files with 2828 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -375,6 +375,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)**  (from HuggingFace).
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
--- a/README_ko.md
+++ b/README_ko.md
@@ -325,6 +325,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
--- a/README_zh-hans.md
+++ b/README_zh-hans.md
@@ -349,6 +349,7 @@ conda install -c huggingface transformers
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (来自 Google AI) 伴随论文 [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) 由 Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu 发布。
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (来自 Google AI) 伴随论文 [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) 由 Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos 发布。
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (来自 Microsoft Research) 伴随论文 [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) 由 Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou 发布。
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (来自 Google/CMU) 伴随论文 [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) 由 Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov 发布。
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (来自 Microsoft) 伴随论文 [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) 由 Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei 发布。
--- a/README_zh-hant.md
+++ b/README_zh-hant.md
@@ -361,6 +361,7 @@ conda install -c huggingface transformers
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released with the paper [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft) released with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -498,6 +498,11 @@
      - local: model_doc/trajectory_transformer
        title: Trajectory Transformer
      title: Reinforcement learning models
+    - isExpanded: false
+      sections:
+      - local: model_doc/time_series_transformer
+        title: Time Series Transformer
+      title: Time series models
    title: Models
  - sections:
    - local: internal/modeling_utils
--- a/docs/source/en/index.mdx
+++ b/docs/source/en/index.mdx
@@ -165,6 +165,7 @@ The documentation is organized into five sections:
 1. **[T5v1.1](model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
+1. **[Time Series Transformer](model_doc/time_series_transformer)**  (from HuggingFace).
 1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
@@ -310,6 +311,7 @@ Flax), PyTorch, and/or TensorFlow.
 |     Swin Transformer V2     |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             T5              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |            TAPAS            |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|   Time Series Transformer   |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |   Trajectory Transformer    |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |       Transformer-XL        |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            TrOCR            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
--- a/docs/source/en/model_doc/time_series_transformer.mdx
+++ b/docs/source/en/model_doc/time_series_transformer.mdx
@@ -0,0 +1,73 @@
+<!--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.
+-->
+
+# Time Series Transformer
+
+<Tip>
+
+This is a recently introduced model so the API hasn't been tested extensively. There may be some bugs or slight
+breaking changes to fix it in the future. If you see something strange, file a [Github Issue](https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title).
+
+</Tip>
+
+## Overview
+
+The Time Series Transformer model is a vanilla encoder-decoder Transformer for time series forecasting.
+
+Tips:
+
+- Similar to other models in the library, [`TimeSeriesTransformerModel`] is the raw Transformer without any head on top, and [`TimeSeriesTransformerForPrediction`]
+adds a distribution head on top of the former, which can be used for time-series forecasting. Note that this is a so-called probabilistic forecasting model, not a
+point forecasting model. This means that the model learns a distribution, from which one can sample. The model doesn't directly output values.
+- [`TimeSeriesTransformerForPrediction`] consists of 2 blocks: an encoder, which takes a `context_length` of time series values as input (called `past_values`),
+and a decoder, which predicts a `prediction_length` of time series values into the future (called `future_values`). During training, one needs to provide
+pairs of (`past_values` and `future_values`) to the model.
+- In addition to the raw (`past_values` and `future_values`), one typically provides additional features to the model. These can be the following:
+    - `past_time_features`: temporal features which the model will add to `past_values`. These serve as "positional encodings" for the Transformer encoder.
+    Examples are "day of the month", "month of the year", etc. as scalar values (and then stacked together as a vector).
+    e.g. if a given time-series value was obtained on the 11th of August, then one could have [11, 8] as time feature vector (11 being "day of the month", 8 being "month of the year").
+    - `future_time_features`: temporal features which the model will add to `future_values`. These serve as "positional encodings" for the Transformer decoder.
+    Examples are "day of the month", "month of the year", etc. as scalar values (and then stacked together as a vector).
+    e.g. if a given time-series value was obtained on the 11th of August, then one could have [11, 8] as time feature vector (11 being "day of the month", 8 being "month of the year").
+    - `static_categorical_features`: categorical features which are static over time (i.e., have the same value for all `past_values` and `future_values`).
+    An example here is the store ID or region ID that identifies a given time-series.
+    Note that these features need to be known for ALL data points (also those in the future).
+    - `static_real_features`: real-valued features which are static over time (i.e., have the same value for all `past_values` and `future_values`).
+    An example here is the image representation of the product for which you have the time-series values (like the [ResNet](resnet) embedding of a "shoe" picture,
+    if your time-series is about the sales of shoes).
+    Note that these features need to be known for ALL data points (also those in the future).
+- The model is trained using "teacher-forcing", similar to how a Transformer is trained for machine translation. This means that, during training, one shifts the
+`future_values` one position to the right as input to the decoder, prepended by the last value of `past_values`. At each time step, the model needs to predict the
+next target. So the set-up of training is similar to a GPT model for language, except that there's no notion of `decoder_start_token_id` (we just use the last value
+of the context as initial input for the decoder).
+- At inference time, we give the final value of the `past_values` as input to the decoder. Next, we can sample from the model to make a prediction at the next time step,
+which is then fed to the decoder in order to make the next prediction (also called autoregressive generation).
+
+
+This model was contributed by [kashif](<https://huggingface.co/kashif).
+
+
+## TimeSeriesTransformerConfig
+
+[[autodoc]] TimeSeriesTransformerConfig
+
+
+## TimeSeriesTransformerModel
+
+[[autodoc]] TimeSeriesTransformerModel
+    - forward
+
+
+## TimeSeriesTransformerForPrediction
+
+[[autodoc]] TimeSeriesTransformerForPrediction
+    - forward
--- a/setup.py
+++ b/setup.py
@@ -281,6 +281,7 @@ extras["vision"] = deps_list("Pillow")
 extras["timm"] = deps_list("timm")
 extras["codecarbon"] = deps_list("codecarbon")

+
 extras["sentencepiece"] = deps_list("sentencepiece", "protobuf")
 extras["testing"] = (
    deps_list(
--- a/src/transformers/init.py
+++ b/src/transformers/init.py
@@ -336,6 +336,10 @@ _import_structure = {
    "models.t5": ["T5_PRETRAINED_CONFIG_ARCHIVE_MAP", "T5Config"],
    "models.tapas": ["TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP", "TapasConfig", "TapasTokenizer"],
    "models.tapex": ["TapexTokenizer"],
+    "models.time_series_transformer": [
+        "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "TimeSeriesTransformerConfig",
+    ],
    "models.trajectory_transformer": [
        "TRAJECTORY_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
        "TrajectoryTransformerConfig",
@@ -815,6 +819,15 @@ else:
    _import_structure["modeling_utils"] = ["PreTrainedModel"]

    # PyTorch models structure
+
+    _import_structure["models.time_series_transformer"].extend(
+        [
+            "TIME_SERIES_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "TimeSeriesTransformerForPrediction",
+            "TimeSeriesTransformerModel",
+            "TimeSeriesTransformerPreTrainedModel",
+        ]
+    )
    _import_structure["models.albert"].extend(
        [
            "ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3286,6 +3299,10 @@ if TYPE_CHECKING:
    from .models.t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
    from .models.tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig, TapasTokenizer
    from .models.tapex import TapexTokenizer
+    from .models.time_series_transformer import (
+        TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        TimeSeriesTransformerConfig,
+    )
    from .models.trajectory_transformer import (
        TRAJECTORY_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
        TrajectoryTransformerConfig,
@@ -4577,6 +4594,12 @@ if TYPE_CHECKING:
            T5PreTrainedModel,
            load_tf_weights_in_t5,
        )
+        from .models.time_series_transformer import (
+            TIME_SERIES_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TimeSeriesTransformerForPrediction,
+            TimeSeriesTransformerModel,
+            TimeSeriesTransformerPreTrainedModel,
+        )
        from .models.trajectory_transformer import (
            TRAJECTORY_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
            TrajectoryTransformerModel,
--- a/src/transformers/models/init.py
+++ b/src/transformers/models/init.py
@@ -139,6 +139,7 @@ from . import (
    t5,
    tapas,
    tapex,
+    time_series_transformer,
    trajectory_transformer,
    transfo_xl,
    trocr,
--- a/src/transformers/models/auto/configuration_auto.py
+++ b/src/transformers/models/auto/configuration_auto.py
@@ -134,6 +134,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
        ("swinv2", "Swinv2Config"),
        ("t5", "T5Config"),
        ("tapas", "TapasConfig"),
+        ("time_series_transformer", "TimeSeriesTransformerConfig"),
        ("trajectory_transformer", "TrajectoryTransformerConfig"),
        ("transfo-xl", "TransfoXLConfig"),
        ("trocr", "TrOCRConfig"),
@@ -262,6 +263,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
        ("swinv2", "SWINV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("t5", "T5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("tapas", "TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("time_series_transformer", "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("transfo-xl", "TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("unispeech", "UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("unispeech-sat", "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -412,6 +414,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
        ("t5v1.1", "T5v1.1"),
        ("tapas", "TAPAS"),
        ("tapex", "TAPEX"),
+        ("time_series_transformer", "Time Series Transformer"),
        ("trajectory_transformer", "Trajectory Transformer"),
        ("transfo-xl", "Transformer-XL"),
        ("trocr", "TrOCR"),
--- a/src/transformers/models/auto/modeling_auto.py
+++ b/src/transformers/models/auto/modeling_auto.py
@@ -130,6 +130,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
        ("swinv2", "Swinv2Model"),
        ("t5", "T5Model"),
        ("tapas", "TapasModel"),
+        ("time_series_transformer", "TimeSeriesTransformerModel"),
        ("trajectory_transformer", "TrajectoryTransformerModel"),
        ("transfo-xl", "TransfoXLModel"),
        ("unispeech", "UniSpeechModel"),
--- a/src/transformers/models/time_series_transformer/init.py
+++ b/src/transformers/models/time_series_transformer/init.py
@@ -0,0 +1,67 @@
+# 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
+
+# rely on isort to merge the imports
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_time_series_transformer": [
+        "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "TimeSeriesTransformerConfig",
+    ],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_time_series_transformer"] = [
+        "TIME_SERIES_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TimeSeriesTransformerForPrediction",
+        "TimeSeriesTransformerModel",
+        "TimeSeriesTransformerPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_time_series_transformer import (
+        TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        TimeSeriesTransformerConfig,
+    )
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_time_series_transformer import (
+            TIME_SERIES_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TimeSeriesTransformerForPrediction,
+            TimeSeriesTransformerModel,
+            TimeSeriesTransformerPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
--- a/src/transformers/models/time_series_transformer/configuration_time_series_transformer.py
+++ b/src/transformers/models/time_series_transformer/configuration_time_series_transformer.py
@@ -0,0 +1,229 @@
+# coding=utf-8
+# Copyright 2022 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.
+""" Time Series Transformer model configuration"""
+
+from typing import List, Optional
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "huggingface/time-series-transformer-tourism-monthly": (
+        "https://huggingface.co/huggingface/time-series-transformer-tourism-monthly/resolve/main/config.json"
+    ),
+    # See all TimeSeriesTransformer models at https://huggingface.co/models?filter=time_series_transformer
+}
+
+
+class TimeSeriesTransformerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`TimeSeriesTransformerModel`]. It is used to
+    instantiate a Time Series Transformer 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 Time Series
+    Transformer
+    [huggingface/time-series-transformer-tourism-monthly](https://huggingface.co/huggingface/time-series-transformer-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.
+        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
+            multivarate targets.
+        scaling (`bool`, *optional* defaults to `True`):
+            Whether to scale the input targets.
+        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. Default is `[1, 2, 3, 4,
+            5, 6, 7]`.
+        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.
+        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.
+
+        Example:
+
+    ```python
+    >>> from transformers import TimeSeriesTransformerConfig, TimeSeriesTransformerModel
+
+    >>> # Initializing a default Time Series Transformer configuration
+    >>> configuration = TimeSeriesTransformerConfig()
+
+    >>> # Randomly initializing a model from the configuration
+    >>> model = TimeSeriesTransformerModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "time_series_transformer"
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+        "num_hidden_layers": "encoder_layers",
+    }
+
+    def __init__(
+        self,
+        input_size: int = 1,
+        prediction_length: Optional[int] = None,
+        context_length: Optional[int] = None,
+        distribution_output: str = "student_t",
+        loss: str = "nll",
+        lags_sequence: List[int] = [1, 2, 3, 4, 5, 6, 7],
+        scaling: bool = True,
+        num_dynamic_real_features: int = 0,
+        num_static_categorical_features: int = 0,
+        num_static_real_features: int = 0,
+        num_time_features: int = 0,
+        cardinality: Optional[List[int]] = None,
+        embedding_dimension: Optional[List[int]] = None,
+        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 = 0.1,
+        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,
+        **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
+        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
+        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 = [1]
+        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.d_model = input_size * len(lags_sequence) + self._number_of_features
+        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
+
+        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
+            + max(1, self.num_static_real_features)  # there is at least one dummy static real feature
+            + 1  # the log(scale)
+        )
--- a/src/transformers/models/time_series_transformer/modeling_time_series_transformer.py
+++ b/src/transformers/models/time_series_transformer/modeling_time_series_transformer.py
--- a/src/transformers/utils/dummy_pt_objects.py
+++ b/src/transformers/utils/dummy_pt_objects.py
@@ -4811,6 +4811,30 @@ def load_tf_weights_in_t5(*args, **kwargs):
    requires_backends(load_tf_weights_in_t5, ["torch"])


+TIME_SERIES_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TimeSeriesTransformerForPrediction(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TimeSeriesTransformerModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TimeSeriesTransformerPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 TRAJECTORY_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None


--- a/tests/models/time_series_transformer/init.py
+++ b/tests/models/time_series_transformer/init.py
--- a/tests/models/time_series_transformer/test_modeling_time_series_transformer.py
+++ b/tests/models/time_series_transformer/test_modeling_time_series_transformer.py
@@ -0,0 +1,438 @@
+# coding=utf-8
+# Copyright 2022 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 TimeSeriesTransformer model. """
+
+import inspect
+import tempfile
+import unittest
+
+from huggingface_hub import hf_hub_download
+from transformers import is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+TOLERANCE = 1e-4
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        TimeSeriesTransformerConfig,
+        TimeSeriesTransformerForPrediction,
+        TimeSeriesTransformerModel,
+    )
+    from transformers.models.time_series_transformer.modeling_time_series_transformer import (
+        TimeSeriesTransformerDecoder,
+        TimeSeriesTransformerEncoder,
+    )
+
+
+@require_torch
+class TimeSeriesTransformerModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        prediction_length=7,
+        context_length=14,
+        cardinality=19,
+        embedding_dimension=5,
+        num_time_features=4,
+        is_training=True,
+        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,
+        lags_sequence=[1, 2, 3, 4, 5],
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.prediction_length = prediction_length
+        self.context_length = context_length
+        self.cardinality = cardinality
+        self.num_time_features = num_time_features
+        self.lags_sequence = lags_sequence
+        self.embedding_dimension = embedding_dimension
+        self.is_training = is_training
+        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.encoder_seq_length = context_length
+        self.decoder_seq_length = prediction_length
+
+    def get_config(self):
+        return TimeSeriesTransformerConfig(
+            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,
+            prediction_length=self.prediction_length,
+            context_length=self.context_length,
+            lags_sequence=self.lags_sequence,
+            num_time_features=self.num_time_features,
+            num_static_categorical_features=1,
+            cardinality=[self.cardinality],
+            embedding_dimension=[self.embedding_dimension],
+        )
+
+    def prepare_time_series_transformer_inputs_dict(self, config):
+        _past_length = config.context_length + max(config.lags_sequence)
+
+        static_categorical_features = ids_tensor([self.batch_size, 1], config.cardinality[0])
+        static_real_features = floats_tensor([self.batch_size, 1])
+
+        past_time_features = floats_tensor([self.batch_size, _past_length, config.num_time_features])
+        past_values = floats_tensor([self.batch_size, _past_length])
+        past_observed_mask = floats_tensor([self.batch_size, _past_length])
+
+        # decoder inputs
+        future_time_features = floats_tensor([self.batch_size, config.prediction_length, config.num_time_features])
+        future_values = floats_tensor([self.batch_size, config.prediction_length])
+
+        inputs_dict = {
+            "past_values": past_values,
+            "static_categorical_features": static_categorical_features,
+            "static_real_features": static_real_features,
+            "past_time_features": past_time_features,
+            "past_observed_mask": past_observed_mask,
+            "future_time_features": future_time_features,
+            "future_values": future_values,
+        }
+        return inputs_dict
+
+    def prepare_config_and_inputs(self):
+        config = self.get_config()
+        inputs_dict = self.prepare_time_series_transformer_inputs_dict(config)
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_common(self):
+        config, inputs_dict = self.prepare_config_and_inputs()
+        return config, inputs_dict
+
+    def check_encoder_decoder_model_standalone(self, config, inputs_dict):
+        model = TimeSeriesTransformerModel(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 = TimeSeriesTransformerEncoder.from_pretrained(tmpdirname).to(torch_device)
+
+        transformer_inputs, _, _ = model.create_network_inputs(**inputs_dict)
+        enc_input = transformer_inputs[:, : config.context_length, ...]
+        dec_input = transformer_inputs[:, config.context_length :, ...]
+
+        encoder_last_hidden_state_2 = encoder(inputs_embeds=enc_input)[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 = TimeSeriesTransformerDecoder.from_pretrained(tmpdirname).to(torch_device)
+
+        last_hidden_state_2 = decoder(
+            inputs_embeds=dec_input,
+            encoder_hidden_states=encoder_last_hidden_state,
+        )[0]
+
+        self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
+
+
+@require_torch
+class TimeSeriesTransformerModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (TimeSeriesTransformerModel, TimeSeriesTransformerForPrediction) if is_torch_available() else ()
+    )
+    all_generative_model_classes = (TimeSeriesTransformerForPrediction,) if is_torch_available() else ()
+    is_encoder_decoder = True
+    test_pruning = False
+    test_head_masking = False
+    test_missing_keys = False
+    test_torchscript = False
+    test_inputs_embeds = False
+    test_model_common_attributes = False
+
+    def setUp(self):
+        self.model_tester = TimeSeriesTransformerModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=TimeSeriesTransformerConfig, has_text_modality=False)
+
+    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_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)
+
+    # Ignore since we have no tokens embeddings
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    # # Input is 'static_categorical_features' not 'input_ids'
+    def test_model_main_input_name(self):
+        model_signature = inspect.signature(getattr(TimeSeriesTransformerModel, "forward"))
+        # The main input is the name of the argument after `self`
+        observed_main_input_name = list(model_signature.parameters.keys())[1]
+        self.assertEqual(TimeSeriesTransformerModel.main_input_name, observed_main_input_name)
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = [
+                "past_values",
+                "past_time_features",
+                "past_observed_mask",
+                "static_categorical_features",
+                "static_real_features",
+                "future_values",
+                "future_time_features",
+            ]
+
+            expected_arg_names.extend(
+                [
+                    "future_observed_mask",
+                    "decoder_attention_mask",
+                    "head_mask",
+                    "decoder_head_mask",
+                    "cross_attn_head_mask",
+                    "encoder_outputs",
+                    "past_key_values",
+                    "output_hidden_states",
+                    "output_attentions",
+                    "use_cache",
+                    "return_dict",
+                ]
+                if "future_observed_mask" in arg_names
+                else [
+                    "decoder_attention_mask",
+                    "head_mask",
+                    "decoder_head_mask",
+                    "cross_attn_head_mask",
+                    "encoder_outputs",
+                    "past_key_values",
+                    "output_hidden_states",
+                    "output_attentions",
+                    "use_cache",
+                    "return_dict",
+                ]
+            )
+
+            self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        seq_len = getattr(self.model_tester, "seq_length", None)
+        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_seq_length],
+            )
+            out_len = len(outputs)
+
+            correct_outlen = 6
+
+            if "last_hidden_state" in outputs:
+                correct_outlen += 1
+
+            if "past_key_values" in outputs:
+                correct_outlen += 1  # past_key_values have been returned
+
+            if "loss" in outputs:
+                correct_outlen += 1
+
+            if "params" in outputs:
+                correct_outlen += 1
+
+            self.assertEqual(out_len, correct_outlen)
+
+            # decoder attentions
+            decoder_attentions = outputs.decoder_attentions
+            self.assertIsInstance(decoder_attentions, (list, tuple))
+            self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(decoder_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, decoder_seq_length, decoder_seq_length],
+            )
+
+            # cross attentions
+            cross_attentions = outputs.cross_attentions
+            self.assertIsInstance(cross_attentions, (list, tuple))
+            self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(cross_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_attention_heads,
+                    decoder_seq_length,
+                    encoder_seq_length,
+                ],
+            )
+
+        # Check attention is always last and order is fine
+        inputs_dict["output_attentions"] = True
+        inputs_dict["output_hidden_states"] = True
+        model = model_class(config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+        self.assertEqual(out_len + 2, len(outputs))
+
+        self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+
+        self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+        self.assertListEqual(
+            list(self_attentions[0].shape[-3:]),
+            [self.model_tester.num_attention_heads, encoder_seq_length, encoder_seq_length],
+        )
+
+
+def prepare_batch(filename="train-batch.pt"):
+    file = hf_hub_download(repo_id="kashif/tourism-monthly-batch", filename=filename, repo_type="dataset")
+    batch = torch.load(file, map_location=torch_device)
+    return batch
+
+
+@require_torch
+@slow
+class TimeSeriesTransformerModelIntegrationTests(unittest.TestCase):
+    def test_inference_no_head(self):
+        model = TimeSeriesTransformerModel.from_pretrained("huggingface/time-series-transformer-tourism-monthly").to(
+            torch_device
+        )
+        batch = prepare_batch()
+
+        with torch.no_grad():
+            output = model(
+                past_values=batch["past_values"],
+                past_time_features=batch["past_time_features"],
+                past_observed_mask=batch["past_observed_mask"],
+                static_categorical_features=batch["static_categorical_features"],
+                static_real_features=batch["static_real_features"],
+                future_values=batch["future_values"],
+                future_time_features=batch["future_time_features"],
+            )[0]
+
+        expected_shape = torch.Size((64, model.config.prediction_length, model.config.d_model))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[-0.3125, -1.2884, -1.1118], [-0.5801, -1.4907, -0.7782], [0.0849, -1.6557, -0.9755]], device=torch_device
+        )
+        self.assertTrue(torch.allclose(output[0, :3, :3], expected_slice, atol=TOLERANCE))
+
+    def test_inference_head(self):
+        model = TimeSeriesTransformerForPrediction.from_pretrained(
+            "huggingface/time-series-transformer-tourism-monthly"
+        ).to(torch_device)
+        batch = prepare_batch("val-batch.pt")
+        with torch.no_grad():
+            output = model(
+                past_values=batch["past_values"],
+                past_time_features=batch["past_time_features"],
+                past_observed_mask=batch["past_observed_mask"],
+                static_categorical_features=batch["static_categorical_features"],
+                static_real_features=batch["static_real_features"],
+                future_time_features=batch["future_time_features"],
+            )[1]
+        expected_shape = torch.Size((64, model.config.prediction_length, model.config.d_model))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[0.9127, -0.2056, -0.5259], [1.0572, 1.4104, -0.1964], [0.1358, 2.0348, 0.5739]], device=torch_device
+        )
+        self.assertTrue(torch.allclose(output[0, :3, :3], expected_slice, atol=TOLERANCE))
+
+    def test_seq_to_seq_generation(self):
+        model = TimeSeriesTransformerForPrediction.from_pretrained(
+            "huggingface/time-series-transformer-tourism-monthly"
+        ).to(torch_device)
+        batch = prepare_batch("val-batch.pt")
+        with torch.no_grad():
+            outputs = model.generate(
+                static_categorical_features=batch["static_categorical_features"],
+                static_real_features=batch["static_real_features"],
+                past_time_features=batch["past_time_features"],
+                past_values=batch["past_values"],
+                future_time_features=batch["future_time_features"],
+                past_observed_mask=batch["past_observed_mask"],
+            )
+        expected_shape = torch.Size((64, model.config.num_parallel_samples, model.config.prediction_length))
+        self.assertEqual(outputs.sequences.shape, expected_shape)
+
+        expected_slice = torch.tensor([2289.5203, 2778.3054, 4648.1313], device=torch_device)
+        mean_prediction = outputs.sequences.mean(dim=1)
+        self.assertTrue(torch.allclose(mean_prediction[0, -3:], expected_slice, rtol=1e-1))
--- a/utils/check_repo.py
+++ b/utils/check_repo.py
@@ -46,6 +46,8 @@ PRIVATE_MODELS = [
 # Being in this list is an exception and should **not** be the rule.
 IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
    # models to ignore for not tested
+    "TimeSeriesTransformerEncoder",  # Building part of bigger (tested) model.
+    "TimeSeriesTransformerDecoder",  # Building part of bigger (tested) model.
    "DeformableDetrEncoder",  # Building part of bigger (tested) model.
    "DeformableDetrDecoder",  # Building part of bigger (tested) model.
    "OPTDecoder",  # Building part of bigger (tested) model.
@@ -132,6 +134,7 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
    # models to ignore for model xxx mapping
+    "TimeSeriesTransformerForPrediction",
    "PegasusXEncoder",
    "PegasusXDecoder",
    "PegasusXDecoderWrapper",