**ThesparsemodelNLLB-MoE(MixtureofExpert)isnowavailable! More details [here](nllb-moe)**
Thismodelwascontributedby[Lysandre](https://huggingface.co/lysandre).Theauthors' code can be found [here](https://github.com/facebookresearch/fairseq/tree/nllb).
## Generating with NLLB
While generating the target text set the `forced_bos_token_id` to the target language id. The following
example shows how to translate English to French using the *facebook/nllb-200-distilled-600M* model.
Note that we'reusingtheBCP-47codeforFrench`fra_Latn`.See[here](https://github.com/facebookresearch/flores/blob/main/flores200/README.md#languages-in-flores-200)
<!--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.
-->
# Nyströmformer
## Overview
The Nyströmformer model was proposed in [*Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention*](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn
Fung, Yin Li, and Vikas Singh.
The abstract from the paper is the following:
*Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component
that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or
dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the
input sequence length has limited its application to longer sequences -- a topic being actively studied in the
community. To address this limitation, we propose Nyströmformer -- a model that exhibits favorable scalability as a
function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention
with O(n) complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of
tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard
sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than
standard self-attention. On longer sequence tasks in the Long Range Arena (LRA) benchmark, Nyströmformer performs
favorably relative to other efficient self-attention methods. Our code is available at this https URL.*
This model was contributed by [novice03](https://huggingface.co/novice03). The original code can be found [here](https://github.com/mlpen/Nystromformer).
<!--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.
-->
# OneFormer
## Overview
The OneFormer model was proposed in [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi. OneFormer is a universal image segmentation framework that can be trained on a single panoptic dataset to perform semantic, instance, and panoptic segmentation tasks. OneFormer uses a task token to condition the model on the task in focus, making the architecture task-guided for training, and task-dynamic for inference.
*Universal Image Segmentation is not a new concept. Past attempts to unify image segmentation in the last decades include scene parsing, panoptic segmentation, and, more recently, new panoptic architectures. However, such panoptic architectures do not truly unify image segmentation because they need to be trained individually on the semantic, instance, or panoptic segmentation to achieve the best performance. Ideally, a truly universal framework should be trained only once and achieve SOTA performance across all three image segmentation tasks. To that end, we propose OneFormer, a universal image segmentation framework that unifies segmentation with a multi-task train-once design. We first propose a task-conditioned joint training strategy that enables training on ground truths of each domain (semantic, instance, and panoptic segmentation) within a single multi-task training process. Secondly, we introduce a task token to condition our model on the task at hand, making our model task-dynamic to support multi-task training and inference. Thirdly, we propose using a query-text contrastive loss during training to establish better inter-task and inter-class distinctions. Notably, our single OneFormer model outperforms specialized Mask2Former models across all three segmentation tasks on ADE20k, CityScapes, and COCO, despite the latter being trained on each of the three tasks individually with three times the resources. With new ConvNeXt and DiNAT backbones, we observe even more performance improvement. We believe OneFormer is a significant step towards making image segmentation more universal and accessible.*
Tips:
- OneFormer requires two inputs during inference: *image* and *task token*.
- During training, OneFormer only uses panoptic annotations.
- If you want to train the model in a distributed environment across multiple nodes, then one should update the
`get_num_masks` function inside in the `OneFormerLoss` class of `modeling_oneformer.py`. When training on multiple nodes, this should be
set to the average number of target masks across all nodes, as can be seen in the original implementation [here](https://github.com/SHI-Labs/OneFormer/blob/33ebb56ed34f970a30ae103e786c0cb64c653d9a/oneformer/modeling/criterion.py#L287).
- One can use [`OneFormerProcessor`] to prepare input images and task inputs for the model and optional targets for the model. [`OneformerProcessor`] wraps [`OneFormerImageProcessor`] and [`CLIPTokenizer`] into a single instance to both prepare the images and encode the task inputs.
- To get the final segmentation, depending on the task, you can call [`~OneFormerProcessor.post_process_semantic_segmentation`] or [`~OneFormerImageProcessor.post_process_instance_segmentation`] or [`~OneFormerImageProcessor.post_process_panoptic_segmentation`]. All three tasks can be solved using [`OneFormerForUniversalSegmentation`] output, panoptic segmentation accepts an optional `label_ids_to_fuse` argument to fuse instances of the target object/s (e.g. sky) together.
The figure below illustrates the architecture of OneFormer. Taken from the [original paper](https://arxiv.org/abs/2211.06220).
This model was contributed by [Jitesh Jain](https://huggingface.co/praeclarumjj3). The original code can be found [here](https://github.com/SHI-Labs/OneFormer).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with OneFormer.
- Demo notebooks regarding inference + fine-tuning on custom data can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/OneFormer).
If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we will review it.
The resource should ideally demonstrate something new instead of duplicating an existing resource.
OpenAI GPT model was proposed in [Improving Language Understanding by Generative Pre-Training](https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf)
by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. It's a causal (unidirectional) transformer
pre-trained using language modeling on a large corpus will long range dependencies, the Toronto Book Corpus.
The abstract from the paper is the following:
*Natural language understanding comprises a wide range of diverse tasks such as textual entailment, question answering,
semantic similarity assessment, and document classification. Although large unlabeled text corpora are abundant,
labeled data for learning these specific tasks is scarce, making it challenging for discriminatively trained models to
perform adequately. We demonstrate that large gains on these tasks can be realized by generative pretraining of a
language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task. In
contrast to previous approaches, we make use of task-aware input transformations during fine-tuning to achieve
effective transfer while requiring minimal changes to the model architecture. We demonstrate the effectiveness of our
approach on a wide range of benchmarks for natural language understanding. Our general task-agnostic model outperforms
discriminatively trained models that use architectures specifically crafted for each task, significantly improving upon
the state of the art in 9 out of the 12 tasks studied.*
Tips:
- GPT is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than
the left.
- GPT was trained with a causal language modeling (CLM) objective and is therefore powerful at predicting the next
token in a sequence. Leveraging this feature allows GPT-2 to generate syntactically coherent text as it can be
observed in the *run_generation.py* example script.
[Write With Transformer](https://transformer.huggingface.co/doc/gpt) is a webapp created and hosted by Hugging Face
showcasing the generative capabilities of several models. GPT is one of them.
This model was contributed by [thomwolf](https://huggingface.co/thomwolf). The original code can be found [here](https://github.com/openai/finetune-transformer-lm).
Note:
If you want to reproduce the original tokenization process of the *OpenAI GPT* paper, you will need to install `ftfy`
and `SpaCy`:
```bash
pip install spacy ftfy==4.4.3
python -m spacy download en
```
If you don't install `ftfy` and `SpaCy`, the [`OpenAIGPTTokenizer`] will default to tokenize
using BERT's `BasicTokenizer` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with OpenAI GPT. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTag pipeline="text-classification"/>
- A blog post on [outperforming OpenAI GPT-3 with SetFit for text-classification](https://www.philschmid.de/getting-started-setfit).
- See also: [Text classification task guide](../tasks/sequence_classification)
<PipelineTag pipeline="text-generation"/>
- A blog on how to [Finetune a non-English GPT-2 Model with Hugging Face](https://www.philschmid.de/fine-tune-a-non-english-gpt-2-model-with-huggingface).
- A blog on [How to generate text: using different decoding methods for language generation with Transformers](https://huggingface.co/blog/how-to-generate) with GPT-2.
- A blog on [Training CodeParrot 🦜 from Scratch](https://huggingface.co/blog/codeparrot), a large GPT-2 model.
- A blog on [Faster Text Generation with TensorFlow and XLA](https://huggingface.co/blog/tf-xla-generate) with GPT-2.
- A blog on [How to train a Language Model with Megatron-LM](https://huggingface.co/blog/megatron-training) with a GPT-2 model.
- A notebook on how to [finetune GPT2 to generate lyrics in the style of your favorite artist](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb). 🌎
- A notebook on how to [finetune GPT2 to generate tweets in the style of your favorite Twitter user](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb). 🌎
- [Causal language modeling](https://huggingface.co/course/en/chapter7/6?fw=pt#training-a-causal-language-model-from-scratch) chapter of the 🤗 Hugging Face Course.
- [`OpenAIGPTLMHeadModel`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#gpt-2gpt-and-causal-language-modeling), [text generation example script](https://github.com/huggingface/transformers/blob/main/examples/pytorch/text-generation/run_generation.py) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFOpenAIGPTLMHeadModel`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_clmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- See also: [Causal language modeling task guide](../tasks/language_modeling)
<PipelineTag pipeline="token-classification"/>
- A course material on [Byte-Pair Encoding tokenization](https://huggingface.co/course/en/chapter6/5).
<!--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.
-->
# OPT
## Overview
The OPT model was proposed in [Open Pre-trained Transformer Language Models](https://arxiv.org/pdf/2205.01068) by Meta AI.
OPT is a series of open-sourced large causal language models which perform similar in performance to GPT3.
The abstract from the paper is the following:
*Large language models, which are often trained for hundreds of thousands of compute days, have shown remarkable capabilities for zero- and few-shot learning. Given their computational cost, these models are difficult to replicate without significant capital. For the few that are available through APIs, no access is granted to the full model weights, making them difficult to study. We present Open Pre-trained Transformers (OPT), a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters, which we aim to fully and responsibly share with interested researchers. We show that OPT-175B is comparable to GPT-3, while requiring only 1/7th the carbon footprint to develop. We are also releasing our logbook detailing the infrastructure challenges we faced, along with code for experimenting with all of the released models.*
Tips:
- OPT has the same architecture as [`BartDecoder`].
- Contrary to GPT2, OPT adds the EOS token `</s>` to the beginning of every prompt. **Note**: Make sure to pass `use_fast=False` when loading OPT's tokenizer with [`AutoTokenizer`] to get the correct tokenizer.
This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ), [Younes Belkada](https://huggingface.co/ybelkada), and [Patrick Von Platen](https://huggingface.co/patrickvonplaten).
The original code can be found [here](https://github.com/facebookresearch/metaseq).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with OPT. If you're
interested in submitting a resource to be included here, please feel free to open a Pull Request and we will review it.
The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTag pipeline="text-generation" />
- A notebook on [fine-tuning OPT with PEFT, bitsandbytes, and Transformers](https://colab.research.google.com/drive/1jCkpikz0J2o20FBQmYmAGdiKmJGOMo-o?usp=sharing). 🌎
- A blog post on [decoding strategies with OPT](https://huggingface.co/blog/introducing-csearch#62-example-two---opt).
- [Causal language modeling](https://huggingface.co/course/en/chapter7/6?fw=pt#training-a-causal-language-model-from-scratch) chapter of the 🤗 Hugging Face Course.
- [`OPTForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#gpt-2gpt-and-causal-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFOPTForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_clmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- [`FlaxOPTForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling#causal-language-modeling).
- [`OPTForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification.ipynb).
<PipelineTag pipeline="question-answering" />
- [`OPTForQuestionAnswering`] is supported by this [question answering example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb).
<!--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.
-->
# OWL-ViT
## Overview
The OWL-ViT (short for Vision Transformer for Open-World Localization) was proposed in [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. OWL-ViT is an open-vocabulary object detection network trained on a variety of (image, text) pairs. It can be used to query an image with one or multiple text queries to search for and detect target objects described in text.
The abstract from the paper is the following:
*Combining simple architectures with large-scale pre-training has led to massive improvements in image classification. For object detection, pre-training and scaling approaches are less well established, especially in the long-tailed and open-vocabulary setting, where training data is relatively scarce. In this paper, we propose a strong recipe for transferring image-text models to open-vocabulary object detection. We use a standard Vision Transformer architecture with minimal modifications, contrastive image-text pre-training, and end-to-end detection fine-tuning. Our analysis of the scaling properties of this setup shows that increasing image-level pre-training and model size yield consistent improvements on the downstream detection task. We provide the adaptation strategies and regularizations needed to attain very strong performance on zero-shot text-conditioned and one-shot image-conditioned object detection. Code and models are available on GitHub.*
## Usage
OWL-ViT is a zero-shot text-conditioned object detection model. OWL-ViT uses [CLIP](clip) as its multi-modal backbone, with a ViT-like Transformer to get visual features and a causal language model to get the text features. To use CLIP for detection, OWL-ViT removes the final token pooling layer of the vision model and attaches a lightweight classification and box head to each transformer output token. Open-vocabulary classification is enabled by replacing the fixed classification layer weights with the class-name embeddings obtained from the text model. The authors first train CLIP from scratch and fine-tune it end-to-end with the classification and box heads on standard detection datasets using a bipartite matching loss. One or multiple text queries per image can be used to perform zero-shot text-conditioned object detection.
[`OwlViTFeatureExtractor`] can be used to resize (or rescale) and normalize images for the model and [`CLIPTokenizer`] is used to encode the text. [`OwlViTProcessor`] wraps [`OwlViTFeatureExtractor`] and [`CLIPTokenizer`] into a single instance to both encode the text and prepare the images. The following example shows how to perform object detection using [`OwlViTProcessor`] and [`OwlViTForObjectDetection`].
```python
>>> import requests
>>> from PIL import Image
>>> import torch
>>> from transformers import OwlViTProcessor, OwlViTForObjectDetection
>>> for box, score, label in zip(boxes, scores, labels):
... box = [round(i, 2) for i in box.tolist()]
... if score >= score_threshold:
... print(f"Detected {text[label]} with confidence {round(score.item(), 3)} at location {box}")
Detected a photo of a cat with confidence 0.707 at location [324.97, 20.44, 640.58, 373.29]
Detected a photo of a cat with confidence 0.717 at location [1.46, 55.26, 315.55, 472.17]
```
This model was contributed by [adirik](https://huggingface.co/adirik). The original code can be found [here](https://github.com/google-research/scenic/tree/main/scenic/projects/owl_vit).
**DISCLAIMER:** If you see something strange, file a [Github Issue](https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title)
and assign @patrickvonplaten.
## Overview
The Pegasus model was proposed in [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/pdf/1912.08777.pdf) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu on Dec 18, 2019.
According to the abstract,
- Pegasus' pretraining task is intentionally similar to summarization: important sentences are removed/masked from an
input document and are generated together as one output sequence from the remaining sentences, similar to an
extractive summary.
- Pegasus achieves SOTA summarization performance on all 12 downstream tasks, as measured by ROUGE and human eval.
This model was contributed by [sshleifer](https://huggingface.co/sshleifer). The Authors' code can be found [here](https://github.com/google-research/pegasus).
Tips:
- Sequence-to-sequence model with the same encoder-decoder model architecture as BART. Pegasus is pre-trained jointly on two self-supervised objective functions: Masked Language Modeling (MLM) and a novel summarization specific pretraining objective, called Gap Sentence Generation (GSG).
* MLM: encoder input tokens are randomly replaced by a mask tokens and have to be predicted by the encoder (like in BERT)
* GSG: whole encoder input sentences are replaced by a second mask token and fed to the decoder, but which has a causal mask to hide the future words like a regular auto-regressive transformer decoder.
## Checkpoints
All the [checkpoints](https://huggingface.co/models?search=pegasus) are fine-tuned for summarization, besides
*pegasus-large*, whence the other checkpoints are fine-tuned:
- Each checkpoint is 2.2 GB on disk and 568M parameters.
- FP16 is not supported (help/ideas on this appreciated!).
- Summarizing xsum in fp32 takes about 400ms/sample, with default parameters on a v100 GPU.
- Full replication results and correctly pre-processed data can be found in this [Issue](https://github.com/huggingface/transformers/issues/6844#issue-689259666).
- [Distilled checkpoints](https://huggingface.co/models?search=distill-pegasus) are described in this [paper](https://arxiv.org/abs/2010.13002).
### Examples
- [Script](https://github.com/huggingface/transformers/tree/main/examples/research_projects/seq2seq-distillation/finetune_pegasus_xsum.sh) to fine-tune pegasus
on the XSUM dataset. Data download instructions at [examples/pytorch/summarization/](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization/README.md).
- FP16 is not supported (help/ideas on this appreciated!).
- The adafactor optimizer is recommended for pegasus fine-tuning.
## Implementation Notes
- All models are transformer encoder-decoders with 16 layers in each component.
- The implementation is completely inherited from [`BartForConditionalGeneration`]
- Some key configuration differences:
- static, sinusoidal position embeddings
- the model starts generating with pad_token_id (which has 0 token_embedding) as the prefix.
- more beams are used (`num_beams=8`)
- All pretrained pegasus checkpoints are the same besides three attributes: `tokenizer.model_max_length` (maximum
input size), `max_length` (the maximum number of tokens to generate) and `length_penalty`.
- The code to convert checkpoints trained in the author's [repo](https://github.com/google-research/pegasus) can be
found in `convert_pegasus_tf_to_pytorch.py`.
## Usage Example
```python
>>> from transformers import PegasusForConditionalGeneration, PegasusTokenizer
>>> import torch
>>> src_text = [
... """ PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."""
... ]
... model_name = "google/pegasus-xsum"
... device = "cuda" if torch.cuda.is_available() else "cpu"
<!--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.
-->
# PEGASUS-X
## Overview
The PEGASUS-X model was proposed in [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao and Peter J. Liu.
PEGASUS-X (PEGASUS eXtended) extends the PEGASUS models for long input summarization through additional long input pretraining and using staggered block-local attention with global tokens in the encoder.
The abstract from the paper is the following:
*While large pretrained Transformer models have proven highly capable at tackling natural language tasks, handling long sequence inputs continues to be a significant challenge. One such task is long input summarization, where inputs are longer than the maximum input context of most pretrained models. Through an extensive set of experiments, we investigate what model architectural changes and pretraining paradigms can most efficiently adapt a pretrained Transformer for long input summarization. We find that a staggered, block-local Transformer with global encoder tokens strikes a good balance of performance and efficiency, and that an additional pretraining phase on long sequences meaningfully improves downstream summarization performance. Based on our findings, we introduce PEGASUS-X, an extension of the PEGASUS model with additional long input pretraining to handle inputs of up to 16K tokens. PEGASUS-X achieves strong performance on long input summarization tasks comparable with much larger models while adding few additional parameters and not requiring model parallelism to train.*
Tips:
* PEGASUS-X uses the same tokenizer as PEGASUS.
This model was contributed by [zphang](<https://huggingface.co/zphang). The original code can be found [here](https://github.com/google-research/pegasus).
- Refer to the [blog post](https://huggingface.co/blog/perceiver) if you want to fully understand how the model works and
is implemented in the library. Note that the models available in the library only showcase some examples of what you can do
with the Perceiver. There are many more use cases, including question answering, named-entity recognition, object detection,
audio classification, video classification, etc.
**Note**:
- Perceiver does **not** work with `torch.nn.DataParallel` due to a bug in PyTorch, see [issue #36035](https://github.com/pytorch/pytorch/issues/36035)
<!--Copyright 2020 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.
-->
# PhoBERT
## Overview
The PhoBERT model was proposed in [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92.pdf) by Dat Quoc Nguyen, Anh Tuan Nguyen.
The abstract from the paper is the following:
*We present PhoBERT with two versions, PhoBERT-base and PhoBERT-large, the first public large-scale monolingual
language models pre-trained for Vietnamese. Experimental results show that PhoBERT consistently outperforms the recent
best pre-trained multilingual model XLM-R (Conneau et al., 2020) and improves the state-of-the-art in multiple
Vietnamese-specific NLP tasks including Part-of-speech tagging, Dependency parsing, Named-entity recognition and
Natural language inference.*
Example of use:
```python
>>> import torch
>>> from transformers import AutoModel, AutoTokenizer
This model was contributed by [dqnguyen](https://huggingface.co/dqnguyen). The original code can be found [here](https://github.com/VinAIResearch/PhoBERT).
<!--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.
-->
# Pix2Struct
## Overview
The Pix2Struct model was proposed in [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
The abstract from the paper is the following:
> Visually-situated language is ubiquitous -- sources range from textbooks with diagrams to web pages with images and tables, to mobile apps with buttons and forms. Perhaps due to this diversity, previous work has typically relied on domain-specific recipes with limited sharing of the underlying data, model architectures, and objectives. We present Pix2Struct, a pretrained image-to-text model for purely visual language understanding, which can be finetuned on tasks containing visually-situated language. Pix2Struct is pretrained by learning to parse masked screenshots of web pages into simplified HTML. The web, with its richness of visual elements cleanly reflected in the HTML structure, provides a large source of pretraining data well suited to the diversity of downstream tasks. Intuitively, this objective subsumes common pretraining signals such as OCR, language modeling, image captioning. In addition to the novel pretraining strategy, we introduce a variable-resolution input representation and a more flexible integration of language and vision inputs, where language prompts such as questions are rendered directly on top of the input image. For the first time, we show that a single pretrained model can achieve state-of-the-art results in six out of nine tasks across four domains: documents, illustrations, user interfaces, and natural images.
Tips:
Pix2Struct has been fine tuned on a variety of tasks and datasets, ranging from image captioning, visual question answering (VQA) over different inputs (books, charts, science diagrams), captioning UI components etc. The full list can be found in Table 1 of the paper.
We therefore advise you to use these models for the tasks they have been fine tuned on. For instance, if you want to use Pix2Struct for UI captioning, you should use the model fine tuned on the UI dataset. If you want to use Pix2Struct for image captioning, you should use the model fine tuned on the natural images captioning dataset and so on.
This model was contributed by [ybelkada](https://huggingface.co/ybelkada).
The original code can be found [here](https://github.com/google-research/pix2struct).
<!--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.
-->
# PoolFormer
## Overview
The PoolFormer model was proposed in [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Sea AI Labs. Instead of designing complicated token mixer to achieve SOTA performance, the target of this work is to demonstrate the competence of transformer models largely stem from the general architecture MetaFormer.
The abstract from the paper is the following:
*Transformers have shown great potential in computer vision tasks. A common belief is their attention-based token mixer module contributes most to their competence. However, recent works show the attention-based module in transformers can be replaced by spatial MLPs and the resulted models still perform quite well. Based on this observation, we hypothesize that the general architecture of the transformers, instead of the specific token mixer module, is more essential to the model's performance. To verify this, we deliberately replace the attention module in transformers with an embarrassingly simple spatial pooling operator to conduct only the most basic token mixing. Surprisingly, we observe that the derived model, termed as PoolFormer, achieves competitive performance on multiple computer vision tasks. For example, on ImageNet-1K, PoolFormer achieves 82.1% top-1 accuracy, surpassing well-tuned vision transformer/MLP-like baselines DeiT-B/ResMLP-B24 by 0.3%/1.1% accuracy with 35%/52% fewer parameters and 48%/60% fewer MACs. The effectiveness of PoolFormer verifies our hypothesis and urges us to initiate the concept of "MetaFormer", a general architecture abstracted from transformers without specifying the token mixer. Based on the extensive experiments, we argue that MetaFormer is the key player in achieving superior results for recent transformer and MLP-like models on vision tasks. This work calls for more future research dedicated to improving MetaFormer instead of focusing on the token mixer modules. Additionally, our proposed PoolFormer could serve as a starting baseline for future MetaFormer architecture design.*
The figure below illustrates the architecture of PoolFormer. Taken from the [original paper](https://arxiv.org/abs/2111.11418).
- PoolFormer has a hierarchical architecture, where instead of Attention, a simple Average Pooling layer is present. All checkpoints of the model can be found on the [hub](https://huggingface.co/models?other=poolformer).
- One can use [`PoolFormerImageProcessor`] to prepare images for the model.
- As most models, PoolFormer comes in different sizes, the details of which can be found in the table below.
This model was contributed by [heytanay](https://huggingface.co/heytanay). The original code can be found [here](https://github.com/sail-sg/poolformer).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with PoolFormer.
<PipelineTag pipeline="image-classification"/>
- [`PoolFormerForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
- See also: [Image classification task guide](../tasks/image_classification)
If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
**DISCLAIMER:** If you see something strange, file a [Github Issue](https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title) and assign
@patrickvonplaten
## Overview
The ProphetNet model was proposed in [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training,](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei
Zhang, Ming Zhou on 13 Jan, 2020.
ProphetNet is an encoder-decoder model and can predict n-future tokens for "ngram" language modeling instead of just
the next token.
The abstract from the paper is the following:
*In this paper, we present a new sequence-to-sequence pretraining model called ProphetNet, which introduces a novel
self-supervised objective named future n-gram prediction and the proposed n-stream self-attention mechanism. Instead of
the optimization of one-step ahead prediction in traditional sequence-to-sequence model, the ProphetNet is optimized by
n-step ahead prediction which predicts the next n tokens simultaneously based on previous context tokens at each time
step. The future n-gram prediction explicitly encourages the model to plan for the future tokens and prevent
overfitting on strong local correlations. We pre-train ProphetNet using a base scale dataset (16GB) and a large scale
dataset (160GB) respectively. Then we conduct experiments on CNN/DailyMail, Gigaword, and SQuAD 1.1 benchmarks for
abstractive summarization and question generation tasks. Experimental results show that ProphetNet achieves new
state-of-the-art results on all these datasets compared to the models using the same scale pretraining corpus.*
Tips:
- ProphetNet is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than
the left.
- The model architecture is based on the original Transformer, but replaces the “standard” self-attention mechanism in the decoder by a a main self-attention mechanism and a self and n-stream (predict) self-attention mechanism.
The Authors' code can be found [here](https://github.com/microsoft/ProphetNet).
## Documentation resources
- [Causal language modeling task guide](../tasks/language_modeling)
<!--Copyright 2021 NVIDIA Corporation and 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.
-->
# QDQBERT
## Overview
The QDQBERT model can be referenced in [Integer Quantization for Deep Learning Inference: Principles and Empirical
Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius
Micikevicius.
The abstract from the paper is the following:
*Quantization techniques can reduce the size of Deep Neural Networks and improve inference latency and throughput by
taking advantage of high throughput integer instructions. In this paper we review the mathematical aspects of
quantization parameters and evaluate their choices on a wide range of neural network models for different application
domains, including vision, speech, and language. We focus on quantization techniques that are amenable to acceleration
by processors with high-throughput integer math pipelines. We also present a workflow for 8-bit quantization that is
able to maintain accuracy within 1% of the floating-point baseline on all networks studied, including models that are
more difficult to quantize, such as MobileNets and BERT-large.*
Tips:
- QDQBERT model adds fake quantization operations (pair of QuantizeLinear/DequantizeLinear ops) to (i) linear layer
inputs and weights, (ii) matmul inputs, (iii) residual add inputs, in BERT model.
- QDQBERT requires the dependency of [Pytorch Quantization Toolkit](https://github.com/NVIDIA/TensorRT/tree/master/tools/pytorch-quantization). To install `pip install pytorch-quantization --extra-index-url https://pypi.ngc.nvidia.com`
- QDQBERT model can be loaded from any checkpoint of HuggingFace BERT model (for example *bert-base-uncased*), and
perform Quantization Aware Training/Post Training Quantization.
- A complete example of using QDQBERT model to perform Quatization Aware Training and Post Training Quantization for
SQUAD task can be found at [transformers/examples/research_projects/quantization-qdqbert/](examples/research_projects/quantization-qdqbert/).
This model was contributed by [shangz](https://huggingface.co/shangz).
### Set default quantizers
QDQBERT model adds fake quantization operations (pair of QuantizeLinear/DequantizeLinear ops) to BERT by
`TensorQuantizer` in [Pytorch Quantization Toolkit](https://github.com/NVIDIA/TensorRT/tree/master/tools/pytorch-quantization). `TensorQuantizer` is the module
for quantizing tensors, with `QuantDescriptor` defining how the tensor should be quantized. Refer to [Pytorch
Quantization Toolkit userguide](https://docs.nvidia.com/deeplearning/tensorrt/pytorch-quantization-toolkit/docs/userguide.html) for more details.
Before creating QDQBERT model, one has to set the default `QuantDescriptor` defining default tensor quantizers.
Example:
```python
>>> import pytorch_quantization.nn as quant_nn
>>> from pytorch_quantization.tensor_quant import QuantDescriptor
>>> # The default tensor quantizer is set to use Max calibration method
Retrieval-augmented generation ("RAG") models combine the powers of pretrained dense retrieval (DPR) and
sequence-to-sequence models. RAG models retrieve documents, pass them to a seq2seq model, then marginalize to generate
outputs. The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing
both retrieval and generation to adapt to downstream tasks.
It is based on the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir
Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
The abstract from the paper is the following:
*Large pre-trained language models have been shown to store factual knowledge in their parameters, and achieve
state-of-the-art results when fine-tuned on downstream NLP tasks. However, their ability to access and precisely
manipulate knowledge is still limited, and hence on knowledge-intensive tasks, their performance lags behind
task-specific architectures. Additionally, providing provenance for their decisions and updating their world knowledge
remain open research problems. Pre-trained models with a differentiable access mechanism to explicit nonparametric
memory can overcome this issue, but have so far been only investigated for extractive downstream tasks. We explore a
general-purpose fine-tuning recipe for retrieval-augmented generation (RAG) — models which combine pre-trained
parametric and non-parametric memory for language generation. We introduce RAG models where the parametric memory is a
pre-trained seq2seq model and the non-parametric memory is a dense vector index of Wikipedia, accessed with a
pre-trained neural retriever. We compare two RAG formulations, one which conditions on the same retrieved passages
across the whole generated sequence, the other can use different passages per token. We fine-tune and evaluate our
models on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art on three open domain QA tasks,
outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures. For language generation
tasks, we find that RAG models generate more specific, diverse and factual language than a state-of-the-art
parametric-only seq2seq baseline.*
This model was contributed by [ola13](https://huggingface.co/ola13).
Tips:
- Retrieval-augmented generation (“RAG”) models combine the powers of pretrained dense retrieval (DPR) and Seq2Seq models. RAG models retrieve docs, pass them to a seq2seq model, then marginalize to generate outputs. The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing both retrieval and generation to adapt to downstream tasks.
<!--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.
-->
# REALM
## Overview
The REALM model was proposed in [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang. It's a
retrieval-augmented language model that firstly retrieves documents from a textual knowledge corpus and then
utilizes retrieved documents to process question answering tasks.
The abstract from the paper is the following:
*Language model pre-training has been shown to capture a surprising amount of world knowledge, crucial for NLP tasks
such as question answering. However, this knowledge is stored implicitly in the parameters of a neural network,
requiring ever-larger networks to cover more facts. To capture knowledge in a more modular and interpretable way, we
augment language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend
over documents from a large corpus such as Wikipedia, used during pre-training, fine-tuning and inference. For the
first time, we show how to pre-train such a knowledge retriever in an unsupervised manner, using masked language
modeling as the learning signal and backpropagating through a retrieval step that considers millions of documents. We
demonstrate the effectiveness of Retrieval-Augmented Language Model pre-training (REALM) by fine-tuning on the
challenging task of Open-domain Question Answering (Open-QA). We compare against state-of-the-art models for both
explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous
methods by a significant margin (4-16% absolute accuracy), while also providing qualitative benefits such as
interpretability and modularity.*
This model was contributed by [qqaatw](https://huggingface.co/qqaatw). The original code can be found
**DISCLAIMER:** This model is still a work in progress, if you see something strange, file a [Github Issue](https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title).
## Overview
The Reformer model was proposed in the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451.pdf) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
The abstract from the paper is the following:
*Large Transformer models routinely achieve state-of-the-art results on a number of tasks but training these models can
be prohibitively costly, especially on long sequences. We introduce two techniques to improve the efficiency of
Transformers. For one, we replace dot-product attention by one that uses locality-sensitive hashing, changing its
complexity from O(L^2) to O(Llog(L)), where L is the length of the sequence. Furthermore, we use reversible residual
layers instead of the standard residuals, which allows storing activations only once in the training process instead of
N times, where N is the number of layers. The resulting model, the Reformer, performs on par with Transformer models
while being much more memory-efficient and much faster on long sequences.*
This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten). The Authors' code can be
found [here](https://github.com/google/trax/tree/master/trax/models/reformer).
Tips:
- Reformer does **not** work with *torch.nn.DataParallel* due to a bug in PyTorch, see [issue #36035](https://github.com/pytorch/pytorch/issues/36035).
- Use Axial position encoding (see below for more details). It’s a mechanism to avoid having a huge positional encoding matrix (when the sequence length is very big) by factorizing it into smaller matrices.
- Replace traditional attention by LSH (local-sensitive hashing) attention (see below for more details). It’s a technique to avoid computing the full product query-key in the attention layers.
- Avoid storing the intermediate results of each layer by using reversible transformer layers to obtain them during the backward pass (subtracting the residuals from the input of the next layer gives them back) or recomputing them for results inside a given layer (less efficient than storing them but saves memory).
- Compute the feedforward operations by chunks and not on the whole batch.
## Axial Positional Encodings
Axial Positional Encodings were first implemented in Google's [trax library](https://github.com/google/trax/blob/4d99ad4965bab1deba227539758d59f0df0fef48/trax/layers/research/position_encodings.py#L29)
and developed by the authors of this model's paper. In models that are treating very long input sequences, the
conventional position id encodings store an embedings vector of size \\(d\\) being the `config.hidden_size` for
every position \\(i, \ldots, n_s\\), with \\(n_s\\) being `config.max_embedding_size`. This means that having
a sequence length of \\(n_s = 2^{19} \approx 0.5M\\) and a `config.hidden_size` of \\(d = 2^{10} \approx 1000\\)
would result in a position encoding matrix:
$$X_{i,j}, \text{ with } i \in \left[1,\ldots, d\right] \text{ and } j \in \left[1,\ldots, n_s\right]$$
which alone has over 500M parameters to store. Axial positional encodings factorize \\(X_{i,j}\\) into two matrices:
$$X^{1}_{i,j}, \text{ with } i \in \left[1,\ldots, d^1\right] \text{ and } j \in \left[1,\ldots, n_s^1\right]$$
and
$$X^{2}_{i,j}, \text{ with } i \in \left[1,\ldots, d^2\right] \text{ and } j \in \left[1,\ldots, n_s^2\right]$$
X^{1}_{i, k}, & \text{if }\ i < d^1 \text{ with } k = j \mod n_s^1 \\
X^{2}_{i - d^1, l}, & \text{if } i \ge d^1 \text{ with } l = \lfloor\frac{j}{n_s^1}\rfloor
\end{cases}$$
Intuitively, this means that a position embedding vector \\(x_j \in \mathbb{R}^{d}\\) is now the composition of two
factorized embedding vectors: \\(x^1_{k, l} + x^2_{l, k}\\), where as the `config.max_embedding_size` dimension
\\(j\\) is factorized into \\(k \text{ and } l\\). This design ensures that each position embedding vector
\\(x_j\\) is unique.
Using the above example again, axial position encoding with \\(d^1 = 2^9, d^2 = 2^9, n_s^1 = 2^9, n_s^2 = 2^{10}\\)
can drastically reduced the number of parameters from 500 000 000 to \\(2^{18} + 2^{19} \approx 780 000\\) parameters, this means 85% less memory usage.
In practice, the parameter `config.axial_pos_embds_dim` is set to a tuple \\((d^1, d^2)\\) which sum has to be
equal to `config.hidden_size` and `config.axial_pos_shape` is set to a tuple \\((n_s^1, n_s^2)\\) which
product has to be equal to `config.max_embedding_size`, which during training has to be equal to the *sequence
length* of the `input_ids`.
## LSH Self Attention
In Locality sensitive hashing (LSH) self attention the key and query projection weights are tied. Therefore, the key
query embedding vectors are also tied. LSH self attention uses the locality sensitive hashing mechanism proposed in
[Practical and Optimal LSH for Angular Distance](https://arxiv.org/abs/1509.02897) to assign each of the tied key
query embedding vectors to one of `config.num_buckets` possible buckets. The premise is that the more "similar"
key query embedding vectors (in terms of *cosine similarity*) are to each other, the more likely they are assigned to
the same bucket.
The accuracy of the LSH mechanism can be improved by increasing `config.num_hashes` or directly the argument
`num_hashes` of the forward function so that the output of the LSH self attention better approximates the output
of the "normal" full self attention. The buckets are then sorted and chunked into query key embedding vector chunks
each of length `config.lsh_chunk_length`. For each chunk, the query embedding vectors attend to its key vectors
(which are tied to themselves) and to the key embedding vectors of `config.lsh_num_chunks_before` previous
neighboring chunks and `config.lsh_num_chunks_after` following neighboring chunks.
For more information, see the [original Paper](https://arxiv.org/abs/2001.04451) or this great [blog post](https://www.pragmatic.ml/reformer-deep-dive/).
Note that `config.num_buckets` can also be factorized into a list \\((n_{\text{buckets}}^1,
n_{\text{buckets}}^2)\\). This way instead of assigning the query key embedding vectors to one of \\((1,\ldots,
n_{\text{buckets}})\\) they are assigned to one of \\((1-1,\ldots, n_{\text{buckets}}^1-1, \ldots,
1-n_{\text{buckets}}^2, \ldots, n_{\text{buckets}}^1-n_{\text{buckets}}^2)\\). This is crucial for very long sequences to
save memory.
When training a model from scratch, it is recommended to leave `config.num_buckets=None`, so that depending on the
sequence length a good value for `num_buckets` is calculated on the fly. This value will then automatically be
saved in the config and should be reused for inference.
Using LSH self attention, the memory and time complexity of the query-key matmul operation can be reduced from
\\(\mathcal{O}(n_s \times n_s)\\) to \\(\mathcal{O}(n_s \times \log(n_s))\\), which usually represents the memory
and time bottleneck in a transformer model, with \\(n_s\\) being the sequence length.
## Local Self Attention
Local self attention is essentially a "normal" self attention layer with key, query and value projections, but is
chunked so that in each chunk of length `config.local_chunk_length` the query embedding vectors only attends to
the key embedding vectors in its chunk and to the key embedding vectors of `config.local_num_chunks_before`
previous neighboring chunks and `config.local_num_chunks_after` following neighboring chunks.
Using Local self attention, the memory and time complexity of the query-key matmul operation can be reduced from
\\(\mathcal{O}(n_s \times n_s)\\) to \\(\mathcal{O}(n_s \times \log(n_s))\\), which usually represents the memory
and time bottleneck in a transformer model, with \\(n_s\\) being the sequence length.
## Training
During training, we must ensure that the sequence length is set to a value that can be divided by the least common
multiple of `config.lsh_chunk_length` and `config.local_chunk_length` and that the parameters of the Axial
Positional Encodings are correctly set as described above. Reformer is very memory efficient so that the model can
easily be trained on sequences as long as 64000 tokens.
For training, the [`ReformerModelWithLMHead`] should be used as follows:
```python
input_ids = tokenizer.encode("This is a sentence from the training data", return_tensors="pt")
<!--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.
-->
# RegNet
## Overview
The RegNet model was proposed in [Designing Network Design Spaces](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
The authors design search spaces to perform Neural Architecture Search (NAS). They first start from a high dimensional search space and iteratively reduce the search space by empirically applying constraints based on the best-performing models sampled by the current search space.
The abstract from the paper is the following:
*In this work, we present a new network design paradigm. Our goal is to help advance the understanding of network design and discover design principles that generalize across settings. Instead of focusing on designing individual network instances, we design network design spaces that parametrize populations of networks. The overall process is analogous to classic manual design of networks, but elevated to the design space level. Using our methodology we explore the structure aspect of network design and arrive at a low-dimensional design space consisting of simple, regular networks that we call RegNet. The core insight of the RegNet parametrization is surprisingly simple: widths and depths of good networks can be explained by a quantized linear function. We analyze the RegNet design space and arrive at interesting findings that do not match the current practice of network design. The RegNet design space provides simple and fast networks that work well across a wide range of flop regimes. Under comparable training settings and flops, the RegNet models outperform the popular EfficientNet models while being up to 5x faster on GPUs.*
Tips:
- One can use [`AutoImageProcessor`] to prepare images for the model.
- The huge 10B model from [Self-supervised Pretraining of Visual Features in the Wild](https://arxiv.org/abs/2103.01988), trained on one billion Instagram images, is available on the [hub](https://huggingface.co/facebook/regnet-y-10b-seer)
This model was contributed by [Francesco](https://huggingface.co/Francesco). The TensorFlow version of the model
was contributed by [sayakpaul](https://huggingface.com/sayakpaul) and [ariG23498](https://huggingface.com/ariG23498).
The original code can be found [here](https://github.com/facebookresearch/pycls).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with RegNet.
<PipelineTag pipeline="image-classification"/>
- [`RegNetForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
- See also: [Image classification task guide](../tasks/image_classification)
If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<!--Copyright 2020 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.
-->
# RemBERT
## Overview
The RemBERT model was proposed in [Rethinking Embedding Coupling in Pre-trained Language Models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, Melvin Johnson, Sebastian Ruder.
The abstract from the paper is the following:
*We re-evaluate the standard practice of sharing weights between input and output embeddings in state-of-the-art
pre-trained language models. We show that decoupled embeddings provide increased modeling flexibility, allowing us to
significantly improve the efficiency of parameter allocation in the input embedding of multilingual models. By
reallocating the input embedding parameters in the Transformer layers, we achieve dramatically better performance on
standard natural language understanding tasks with the same number of parameters during fine-tuning. We also show that
allocating additional capacity to the output embedding provides benefits to the model that persist through the
fine-tuning stage even though the output embedding is discarded after pre-training. Our analysis shows that larger
output embeddings prevent the model's last layers from overspecializing to the pre-training task and encourage
Transformer representations to be more general and more transferable to other tasks and languages. Harnessing these
findings, we are able to train models that achieve strong performance on the XTREME benchmark without increasing the
number of parameters at the fine-tuning stage.*
Tips:
For fine-tuning, RemBERT can be thought of as a bigger version of mBERT with an ALBERT-like factorization of the
embedding layer. The embeddings are not tied in pre-training, in contrast with BERT, which enables smaller input
embeddings (preserved during fine-tuning) and bigger output embeddings (discarded at fine-tuning). The tokenizer is
also similar to the Albert one rather than the BERT one.
