Perceiver IO model pre-trained on the Masked Language Modeling (MLM) task proposed in BERT using a large text corpus obtained by combining English Wikipedia and C4. 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 masked language modeling, the output is a tensor containing the prediction scores of the language modeling head, of shape (batch_size, seq_length, vocab_size).

As the time and memory requirements of the self-attention mechanism don't depend on the size of the inputs, the Perceiver IO authors train the model directly on raw UTF-8 bytes, rather than on subwords as is done in models like BERT, RoBERTa and GPT-2. This has many benefits: one doesn't need to train a tokenizer before training the model, one doesn't need to maintain a (fixed) vocabulary file, and this also doesn't hurt model performance as shown by Bostrom et al., 2020.

By pre-training the model, it learns an inner representation of language that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard classifier using the features produced by the Perceiver model as inputs.

You can use the raw model for masked language modeling, but the model is intended to be fine-tuned on a labeled dataset. See the model hub to look for fine-tuned versions on a task that interests you.

Here is how to use this model in PyTorch:

This model was pretrained on a combination of English Wikipedia and C4. 70% of the training tokens were sampled from the C4 dataset and the remaining 30% from Wikipedia. The authors concatenate 10 documents before splitting into crops to reduce wasteful computation on padding tokens.

Text preprocessing is trivial: it only involves encoding text into UTF-8 bytes, and padding them up to the same length (2048).

Hyperparameter details can be found in table 9 of the paper.

This model is able to achieve an average score of 81.8 on GLUE. For more details, we refer to table 3 of the original paper.