<!--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.
-->
# ResNet
## Overview
The ResNet model was proposed in [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. Our implementation follows the small changes made by [Nvidia](https://catalog.ngc.nvidia.com/orgs/nvidia/resources/resnet_50_v1_5_for_pytorch), we apply the `stride=2` for downsampling in bottleneck's `3x3` conv and not in the first `1x1`. This is generally known as "ResNet v1.5".
ResNet introduced residual connections, they allow to train networks with an unseen number of layers (up to 1000). ResNet won the 2015 ILSVRC & COCO competition, one important milestone in deep computer vision.
The abstract from the paper is the following:
*Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers.
The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.*
Tips:
- One can use [`AutoImageProcessor`] to prepare images for the model.
The figure below illustrates the architecture of ResNet. Taken from the [original paper](https://arxiv.org/abs/1512.03385).
This model was contributed by [Francesco](https://huggingface.co/Francesco). The TensorFlow version of this model was added by [amyeroberts](https://huggingface.co/amyeroberts). The original code can be found [here](https://github.com/KaimingHe/deep-residual-networks).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ResNet.
<PipelineTag pipeline="image-classification"/>
- [`ResNetForImageClassification`] 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 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.
-->
# RoBERTa-PreLayerNorm
## Overview
The RoBERTa-PreLayerNorm model was proposed in [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://arxiv.org/abs/1904.01038) by Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, Michael Auli.
It is identical to using the `--encoder-normalize-before` flag in [fairseq](https://fairseq.readthedocs.io/).
The abstract from the paper is the following:
*fairseq is an open-source sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks. The toolkit is based on PyTorch and supports distributed training across multiple GPUs and machines. We also support fast mixed-precision training and inference on modern GPUs.*
Tips:
- The implementation is the same as [Roberta](roberta) except instead of using _Add and Norm_ it does _Norm and Add_. _Add_ and _Norm_ refers to the Addition and LayerNormalization as described in [Attention Is All You Need](https://arxiv.org/abs/1706.03762).
- This is identical to using the `--encoder-normalize-before` flag in [fairseq](https://fairseq.readthedocs.io/).
This model was contributed by [andreasmaden](https://huggingface.co/andreasmaden).
The original code can be found [here](https://github.com/princeton-nlp/DinkyTrain).
Levy,MikeLewis,LukeZettlemoyer,VeselinStoyanov.ItisbasedonGoogle's BERT model released in 2018.
It builds on BERT and modifies key hyperparameters, removing the next-sentence pretraining objective and training with
much larger mini-batches and learning rates.
The abstract from the paper is the following:
*Language model pretraining has led to significant performance gains but careful comparison between different
approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication
study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every
model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise questions about the source of recently
reported improvements. We release our models and code.*
Tips:
- This implementation is the same as [`BertModel`] with a tiny embeddings tweak as well as a setup
for Roberta pretrained models.
- RoBERTa has the same architecture as BERT, but uses a byte-level BPE as a tokenizer (same as GPT-2) and uses a
different pretraining scheme.
- RoBERTa doesn'thave`token_type_ids`,youdon't need to indicate which token belongs to which segment. Just
separate your segments with the separation token `tokenizer.sep_token` (or `</s>`)
- Same as BERT with better pretraining tricks:
* dynamic masking: tokens are masked differently at each epoch, whereas BERT does it once and for all
* together to reach 512 tokens (so the sentences are in an order than may span several documents)
* train with larger batches
* use BPE with bytes as a subunit and not characters (because of unicode characters)
- [CamemBERT](camembert) is a wrapper around RoBERTa. Refer to this page for usage examples.
This model was contributed by [julien-c](https://huggingface.co/julien-c). The original code can be found [here](https://github.com/pytorch/fairseq/tree/master/examples/roberta).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with RoBERTa. If you'reinterestedinsubmittingaresourcetobeincludedhere,pleasefeelfreetoopenaPullRequestandwe'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTag pipeline="text-classification"/>
- A blog on [Getting Started with Sentiment Analysis on Twitter](https://huggingface.co/blog/sentiment-analysis-twitter) using RoBERTa and the [Inference API](https://huggingface.co/inference-api).
- A blog on [Opinion Classification with Kili and Hugging Face AutoTrain](https://huggingface.co/blog/opinion-classification-with-kili) using RoBERTa.
- A notebook on how to [finetune RoBERTa for sentiment analysis](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb). 🌎
- [`RobertaForSequenceClassification`] 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).
- [`TFRobertaForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification-tf.ipynb).
- [`FlaxRobertaForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification_flax.ipynb).
- [`RobertaForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification.ipynb).
- [`TFRobertaForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification-tf.ipynb).
- [`FlaxRobertaForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/token-classification).
- [Token classification](https://huggingface.co/course/chapter7/2?fw=pt) chapter of the 🤗 Hugging Face Course.
- A blog on [How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train) with RoBERTa.
- [`RobertaForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#robertabertdistilbert-and-masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFRobertaForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_mlmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- [`FlaxRobertaForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling#masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/masked_language_modeling_flax.ipynb).
- [Masked language modeling](https://huggingface.co/course/chapter7/3?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Masked language modeling task guide](../tasks/masked_language_modeling)
<PipelineTag pipeline="question-answering"/>
- A blog on [Accelerated Inference with Optimum and Transformers Pipelines](https://huggingface.co/blog/optimum-inference) with RoBERTa for question answering.
- [`RobertaForQuestionAnswering`] is supported by this [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).
- [`TFRobertaForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering-tf.ipynb).
- [`FlaxRobertaForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/flax/question-answering).
- [Question answering](https://huggingface.co/course/chapter7/7?fw=pt) chapter of the 🤗 Hugging Face Course.
- [`RobertaForMultipleChoice`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/multiple-choice) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice.ipynb).
- [`TFRobertaForMultipleChoice`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/multiple-choice) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice-tf.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.
-->
# RoCBert
## Overview
The RoCBert model was proposed in [RoCBert: Robust Chinese Bert with Multimodal Contrastive Pretraining](https://aclanthology.org/2022.acl-long.65.pdf) by HuiSu, WeiweiShi, XiaoyuShen, XiaoZhou, TuoJi, JiaruiFang, JieZhou.
It's a pretrained Chinese language model that is robust under various forms of adversarial attacks.
The abstract from the paper is the following:
*Large-scale pretrained language models have achieved SOTA results on NLP tasks. However, they have been shown
vulnerable to adversarial attacks especially for logographic languages like Chinese. In this work, we propose
ROCBERT: a pretrained Chinese Bert that is robust to various forms of adversarial attacks like word perturbation,
synonyms, typos, etc. It is pretrained with the contrastive learning objective which maximizes the label consistency
under different synthesized adversarial examples. The model takes as input multimodal information including the
semantic, phonetic and visual features. We show all these features are important to the model robustness since the
attack can be performed in all the three forms. Across 5 Chinese NLU tasks, ROCBERT outperforms strong baselines under
three blackbox adversarial algorithms without sacrificing the performance on clean testset. It also performs the best
in the toxic content detection task under human-made attacks.*
This model was contributed by [weiweishi](https://huggingface.co/weiweishi).
<!--Copyright 2021 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.
-->
# RoFormer
## Overview
The RoFormer model was proposed in [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
The abstract from the paper is the following:
*Position encoding in transformer architecture provides supervision for dependency modeling between elements at
different positions in the sequence. We investigate various methods to encode positional information in
transformer-based language models and propose a novel implementation named Rotary Position Embedding(RoPE). The
proposed RoPE encodes absolute positional information with rotation matrix and naturally incorporates explicit relative
position dependency in self-attention formulation. Notably, RoPE comes with valuable properties such as flexibility of
being expand to any sequence lengths, decaying inter-token dependency with increasing relative distances, and
capability of equipping the linear self-attention with relative position encoding. As a result, the enhanced
transformer with rotary position embedding, or RoFormer, achieves superior performance in tasks with long texts. We
release the theoretical analysis along with some preliminary experiment results on Chinese data. The undergoing
experiment for English benchmark will soon be updated.*
Tips:
- RoFormer is a BERT-like autoencoding model with rotary position embeddings. Rotary position embeddings have shown
improved performance on classification tasks with long texts.
This model was contributed by [junnyu](https://huggingface.co/junnyu). The original code can be found [here](https://github.com/ZhuiyiTechnology/roformer).
<!--Copyright 2021 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.
-->
# SegFormer
## Overview
The SegFormer model was proposed in [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping
Luo. The model consists of a hierarchical Transformer encoder and a lightweight all-MLP decode head to achieve great
results on image segmentation benchmarks such as ADE20K and Cityscapes.
The abstract from the paper is the following:
*We present SegFormer, a simple, efficient yet powerful semantic segmentation framework which unifies Transformers with
lightweight multilayer perception (MLP) decoders. SegFormer has two appealing features: 1) SegFormer comprises a novel
hierarchically structured Transformer encoder which outputs multiscale features. It does not need positional encoding,
thereby avoiding the interpolation of positional codes which leads to decreased performance when the testing resolution
differs from training. 2) SegFormer avoids complex decoders. The proposed MLP decoder aggregates information from
different layers, and thus combining both local attention and global attention to render powerful representations. We
show that this simple and lightweight design is the key to efficient segmentation on Transformers. We scale our
approach up to obtain a series of models from SegFormer-B0 to SegFormer-B5, reaching significantly better performance
and efficiency than previous counterparts. For example, SegFormer-B4 achieves 50.3% mIoU on ADE20K with 64M parameters,
being 5x smaller and 2.2% better than the previous best method. Our best model, SegFormer-B5, achieves 84.0% mIoU on
Cityscapes validation set and shows excellent zero-shot robustness on Cityscapes-C.*
The figure below illustrates the architecture of SegFormer. Taken from the [original paper](https://arxiv.org/abs/2105.15203).
This model was contributed by [nielsr](https://huggingface.co/nielsr). The TensorFlow version
of the model was contributed by [sayakpaul](https://huggingface.co/sayakpaul). The original code can be found [here](https://github.com/NVlabs/SegFormer).
Tips:
- SegFormer consists of a hierarchical Transformer encoder, and a lightweight all-MLP decoder head.
[`SegformerModel`] is the hierarchical Transformer encoder (which in the paper is also referred to
as Mix Transformer or MiT). [`SegformerForSemanticSegmentation`] adds the all-MLP decoder head on
top to perform semantic segmentation of images. In addition, there's
[`SegformerForImageClassification`] which can be used to - you guessed it - classify images. The
authors of SegFormer first pre-trained the Transformer encoder on ImageNet-1k to classify images. Next, they throw
away the classification head, and replace it by the all-MLP decode head. Next, they fine-tune the model altogether on
ADE20K, Cityscapes and COCO-stuff, which are important benchmarks for semantic segmentation. All checkpoints can be
found on the [hub](https://huggingface.co/models?other=segformer).
- The quickest way to get started with SegFormer is by checking the [example notebooks](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SegFormer) (which showcase both inference and
fine-tuning on custom data). One can also check out the [blog post](https://huggingface.co/blog/fine-tune-segformer) introducing SegFormer and illustrating how it can be fine-tuned on custom data.
- TensorFlow users should refer to [this repository](https://github.com/deep-diver/segformer-tf-transformers) that shows off-the-shelf inference and fine-tuning.
- One can also check out [this interactive demo on Hugging Face Spaces](https://huggingface.co/spaces/chansung/segformer-tf-transformers)
to try out a SegFormer model on custom images.
- SegFormer works on any input size, as it pads the input to be divisible by `config.patch_sizes`.
- One can use [`SegformerImageProcessor`] to prepare images and corresponding segmentation maps
for the model. Note that this image processor is fairly basic and does not include all data augmentations used in
the original paper. The original preprocessing pipelines (for the ADE20k dataset for instance) can be found [here](https://github.com/NVlabs/SegFormer/blob/master/local_configs/_base_/datasets/ade20k_repeat.py). The most
important preprocessing step is that images and segmentation maps are randomly cropped and padded to the same size,
such as 512x512 or 640x640, after which they are normalized.
- One additional thing to keep in mind is that one can initialize [`SegformerImageProcessor`] with
`reduce_labels` set to `True` or `False`. In some datasets (like ADE20k), the 0 index is used in the annotated
segmentation maps for background. However, ADE20k doesn't include the "background" class in its 150 labels.
Therefore, `reduce_labels` is used to reduce all labels by 1, and to make sure no loss is computed for the
background class (i.e. it replaces 0 in the annotated maps by 255, which is the *ignore_index* of the loss function
used by [`SegformerForSemanticSegmentation`]). However, other datasets use the 0 index as
background class and include this class as part of all labels. In that case, `reduce_labels` should be set to
`False`, as loss should also be computed for the background class.
- As most models, SegFormer comes in different sizes, the details of which can be found in the table below
(taken from Table 7 of the [original paper](https://arxiv.org/abs/2105.15203)).
Note that MiT in the above table refers to the Mix Transformer encoder backbone introduced in SegFormer. For
SegFormer's results on the segmentation datasets like ADE20k, refer to the [paper](https://arxiv.org/abs/2105.15203).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SegFormer.
<PipelineTag pipeline="image-classification"/>
- [`SegformerForImageClassification`] 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).
- [`SegformerForSemanticSegmentation`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/semantic-segmentation).
- A blog on fine-tuning SegFormer on a custom dataset can be found [here](https://huggingface.co/blog/fine-tune-segformer).
- More demo notebooks on SegFormer (both inference + fine-tuning on a custom dataset) can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SegFormer).
- [`TFSegformerForSemanticSegmentation`] is supported by this [example notebook](https://github.com/huggingface/notebooks/blob/main/examples/semantic_segmentation-tf.ipynb).
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 2021 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.
-->
# Speech Encoder Decoder Models
The [`SpeechEncoderDecoderModel`] can be used to initialize a speech-to-text model
with any pretrained speech autoencoding model as the encoder (*e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert)) and any pretrained autoregressive model as the decoder.
The effectiveness of initializing speech-sequence-to-text-sequence models with pretrained checkpoints for speech
recognition and speech translation has *e.g.* been shown in [Large-Scale Self- and Semi-Supervised Learning for Speech
Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli,
Alexis Conneau.
An example of how to use a [`SpeechEncoderDecoderModel`] for inference can be seen in [Speech2Text2](speech_to_text_2).
## Randomly initializing `SpeechEncoderDecoderModel` from model configurations.
[`SpeechEncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`Wav2Vec2Model`] configuration for the encoder
and the default [`BertForCausalLM`] configuration for the decoder.
```python
>>> from transformers import BertConfig, Wav2Vec2Config, SpeechEncoderDecoderConfig, SpeechEncoderDecoderModel
>>> model = SpeechEncoderDecoderModel(config=config)
```
## Initialising `SpeechEncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
[`SpeechEncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained Transformer-based speech model, *e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert) can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
Initializing [`SpeechEncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
To do so, the `SpeechEncoderDecoderModel` class provides a [`SpeechEncoderDecoderModel.from_encoder_decoder_pretrained`] method.
```python
>>> from transformers import SpeechEncoderDecoderModel
>>> model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
## Loading an existing `SpeechEncoderDecoderModel` checkpoint and perform inference.
To load fine-tuned checkpoints of the `SpeechEncoderDecoderModel` class, [`SpeechEncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
```python
>>> from transformers import Wav2Vec2Processor, SpeechEncoderDecoderModel
>>> from datasets import load_dataset
>>> import torch
>>> # load a fine-tuned speech translation model and corresponding processor
>>> model = SpeechEncoderDecoderModel.from_pretrained("facebook/wav2vec2-xls-r-300m-en-to-15")
