Document Image Transformer (DiT) model pre-trained on IIT-CDIP (Lewis et al., 2006), a dataset that includes 42 million document images. 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.
1from transformers import BeitImageProcessor, BeitForMaskedImageModeling
2import torch
3from PIL import Image
4
5image = Image.open('path_to_your_document_image').convert('RGB')
6
7processor = BeitImageProcessor.from_pretrained("microsoft/dit-large")
8model = BeitForMaskedImageModeling.from_pretrained("microsoft/dit-large")
9
10num_patches = (model.config.image_size // model.config.patch_size) ** 2
11pixel_values = processor(images=image, return_tensors="pt").pixel_values
12# create random boolean mask of shape (batch_size, num_patches)
13bool_masked_pos = torch.randint(low=0, high=2, size=(1, num_patches)).bool()
14
15outputs = model(pixel_values, bool_masked_pos=bool_masked_pos)
16loss, logits = outputs.loss, outputs.logits
1@article{Lewis2006BuildingAT,
2 title={Building a test collection for complex document information processing},
3 author={David D. Lewis and Gady Agam and Shlomo Engelson Argamon and Ophir Frieder and David A. Grossman and Jefferson Heard},
4 journal={Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval},
5 year={2006}
6}
