Document Image Transformer (DiT) model pre-trained on IIT-CDIP (Lewis et al., 2006), a dataset that includes 42 million document images and fine-tuned on RVL-CDIP, a dataset consisting of 400,000 grayscale images in 16 classes, with 25,000 images per class. It was introduced in the paper DiT: Self-supervised Pre-training for Document Image Transformer by Li et al. and first released in this repository. Note that DiT is identical to the architecture of BEiT.

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

The Document Image Transformer (DiT) is a transformer encoder model (BERT-like) pre-trained on a large collection of images in a self-supervised fashion. The pre-training objective for the model is to predict visual tokens from the encoder of a discrete VAE (dVAE), based on masked patches.

Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder.

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 document images for instance, you can train a standard classifier by placing a linear layer on top of the pre-trained encoder.

You can use the raw model for encoding document images into a vector space, but it's mostly meant to be fine-tuned on tasks like document image classification, table detection or document layout analysis. 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: