Perceiver IO model pre-trained on ImageNet (14 million images, 1,000 classes) at resolution 224x224. It was introduced in the paper Perceiver IO: A General Architecture for Structured Inputs & Outputs by Jaegle et al. and first released in this repository.

Disclaimer: The team releasing Perceiver IO did not write a model card for this model so this model card has been written by the Hugging Face team.

Perceiver IO is a transformer encoder model that can be applied on any modality (text, images, audio, video, ...). The core idea is to employ the self-attention mechanism on a not-too-large set of latent vectors (e.g. 256 or 512), and only use the inputs to perform cross-attention with the latents. This allows for the time and memory requirements of the self-attention mechanism to not depend on the size of the inputs.

To decode, the authors employ so-called decoder queries, which allow to flexibly decode the final hidden states of the latents to produce outputs of arbitrary size and semantics. For image classification, the output is a tensor containing the logits, of shape (batch_size, num_labels).

As the time and memory requirements of the self-attention mechanism don't depend on the size of the inputs, the Perceiver IO authors can train the model directly on raw pixel values, rather than on patches as is done in ViT. This particular model only adds learned 1D position embeddings to the pixel values, hence it is given no privileged information about the 2D structure of images.

By pre-training the model, it learns an inner representation of images that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled images for instance, you can train a standard classifier by replacing the classification decoder.

You can use the raw model for image classification. See the model hub to look for other fine-tuned versions on a task that may interest you.

Here is how to use this model in PyTorch:

This model was pretrained on ImageNet, a dataset consisting of 14 million images and 1k classes.

Images are center cropped and resized to a resolution of 224x224 and normalized across the RGB channels. Note that data augmentation was used during pre-training, as explained in Appendix H of the paper.

Hyperparameter details can be found in Appendix H of the paper.

This model is able to achieve a top-1 accuracy of 72.7 on ImageNet-1k, despite having no privileged information about the 2D structure of images.