This model was released on 2020-12-23 and added to Hugging Face Transformers on 2021-04-13.

DeiT

Overview

The DeiT model was proposed in Training data-efficient image transformers & distillation through attention by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou. The Vision Transformer (ViT) introduced in Dosovitskiy et al., 2020 has shown that one can match or even outperform existing convolutional neural networks using a Transformer encoder (BERT-like). However, the ViT models introduced in that paper required training on expensive infrastructure for multiple weeks, using external data. DeiT (data-efficient image transformers) are more efficiently trained transformers for image classification, requiring far less data and far less computing resources compared to the original ViT models.

The abstract from the paper is the following:

Recently, neural networks purely based on attention were shown to address image understanding tasks such as image classification. However, these visual transformers are pre-trained with hundreds of millions of images using an expensive infrastructure, thereby limiting their adoption. In this work, we produce a competitive convolution-free transformer by training on Imagenet only. We train them on a single computer in less than 3 days. Our reference vision transformer (86M parameters) achieves top-1 accuracy of 83.1% (single-crop evaluation) on ImageNet with no external data. More importantly, we introduce a teacher-student strategy specific to transformers. It relies on a distillation token ensuring that the student learns from the teacher through attention. We show the interest of this token-based distillation, especially when using a convnet as a teacher. This leads us to report results competitive with convnets for both Imagenet (where we obtain up to 85.2% accuracy) and when transferring to other tasks. We share our code and models.

This model was contributed by nielsr.

Usage tips

• Compared to ViT, DeiT models use a so-called distillation token to effectively learn from a teacher (which, in the DeiT paper, is a ResNet like-model). The distillation token is learned through backpropagation, by interacting with the class ([CLS]) and patch tokens through the self-attention layers.
• There are 2 ways to fine-tune distilled models, either (1) in a classic way, by only placing a prediction head on top of the final hidden state of the class token and not using the distillation signal, or (2) by placing both a prediction head on top of the class token and on top of the distillation token. In that case, the [CLS] prediction head is trained using regular cross-entropy between the prediction of the head and the ground-truth label, while the distillation prediction head is trained using hard distillation (cross-entropy between the prediction of the distillation head and the label predicted by the teacher). At inference time, one takes the average prediction between both heads as final prediction. (2) is also called "fine-tuning with distillation", because one relies on a teacher that has already been fine-tuned on the downstream dataset. In terms of models, (1) corresponds to [DeiTForImageClassification] and (2) corresponds to [DeiTForImageClassificationWithTeacher].
• Note that the authors also did try soft distillation for (2) (in which case the distillation prediction head is trained using KL divergence to match the softmax output of the teacher), but hard distillation gave the best results.
• All released checkpoints were pre-trained and fine-tuned on ImageNet-1k only. No external data was used. This is in contrast with the original ViT model, which used external data like the JFT-300M dataset/Imagenet-21k for pre-training.
• The authors of DeiT also released more efficiently trained ViT models, which you can directly plug into [ViTModel] or [ViTForImageClassification]. Techniques like data augmentation, optimization, and regularization were used in order to simulate training on a much larger dataset (while only using ImageNet-1k for pre-training). There are 4 variants available (in 3 different sizes): facebook/deit-tiny-patch16-224, facebook/deit-small-patch16-224, facebook/deit-base-patch16-224 and facebook/deit-base-patch16-384. Note that one should use [DeiTImageProcessor] in order to prepare images for the model.

Using Scaled Dot Product Attention (SDPA)

PyTorch includes a native scaled dot-product attention (SDPA) operator as part of torch.nn.functional. This function encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the official documentation or the GPU Inference page for more information.

SDPA is used by default for torch>=2.1.1 when an implementation is available, but you may also set attn_implementation="sdpa" in from_pretrained() to explicitly request SDPA to be used.

from transformers import DeiTForImageClassification


model = DeiTForImageClassification.from_pretrained("facebook/deit-base-distilled-patch16-224", attn_implementation="sdpa", device_map="auto")
...

For the best speedups, we recommend loading the model in half-precision (e.g. torch.float16 or torch.bfloat16).

On a local benchmark (A100-40GB, PyTorch 2.3.0, OS Ubuntu 22.04) with float32 and facebook/deit-base-distilled-patch16-224 model, we saw the following speedups during inference.

Resources

A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with DeiT.

• [DeiTForImageClassification] is supported by this example script and notebook.
• See also: Image classification task guide

Besides that:

• [DeiTForMaskedImageModeling] is supported by this example script.

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.

DeiTConfig

[[autodoc]] DeiTConfig

DeiTImageProcessor

[[autodoc]] DeiTImageProcessor - preprocess

DeiTImageProcessorPil

[[autodoc]] DeiTImageProcessorPil - preprocess

DeiTModel

[[autodoc]] DeiTModel - forward

DeiTForMaskedImageModeling

[[autodoc]] DeiTForMaskedImageModeling - forward

DeiTForImageClassification

[[autodoc]] DeiTForImageClassification - forward

DeiTForImageClassificationWithTeacher

[[autodoc]] DeiTForImageClassificationWithTeacher - forward