Sharing models#

docTR is integrated with the Hugging Face Hub: load models shared by the community in one line, and publish your own.

Load a model from the Hub#

from_hub() downloads a model and rebuilds it with its configuration (architecture, vocabulary, classes…). Plug it into any predictor:

from doctr.io import DocumentFile
from doctr.models import from_hub, ocr_predictor

det_model = from_hub("Felix92/doctr-torch-db-mobilenet-v3-large")
reco_model = from_hub("Felix92/doctr-torch-parseq-multilingual-v1")

predictor = ocr_predictor(det_arch=det_model, reco_arch=reco_model)
result = predictor(DocumentFile.from_images("path/to/page.jpg"))

Publish a model#

You need a Hugging Face account and Git LFS installed.

from doctr.models import login_to_hub, push_to_hf_hub, recognition

login_to_hub()  # paste an access token with write permissions

model = recognition.crnn_mobilenet_v3_large(pretrained=True)
push_to_hf_hub(
    model,
    model_name="doctr-crnn-mobilenet-v3-large-french-v1",
    task="recognition",  # classification, detection, recognition, layout or table_structure
    arch="crnn_mobilenet_v3_large",
)

Existing repositories are never overwritten. The reference training scripts can also publish the model at the end of the training:

python references/recognition/train.py crnn_mobilenet_v3_large --name doctr-crnn-mobilenet-v3-large --push-to-hub

Naming conventions#

Task	Suggested name
Classification	`doctr-<architecture>-<vocab>`
Detection	`doctr-<architecture>`
Recognition	`doctr-<architecture>-<vocab>`
Layout	`doctr-<architecture>`
Table structure	`doctr-<architecture>`

Community models#

Models trained and shared by the community. Open a pull request to add yours!

Task	Architecture	Repository	Language
Detection	`db_resnet50`	`rania-sr/doctr-Detection-model-v1-arabic`	Arabic
Recognition	`parseq`	`Felix92/doctr-torch-parseq-multilingual-v1`	Multilingual
Recognition	`parseq`	`rania-sr/doctr-model-v1-arabic`	Arabic
Recognition	`crnn_vgg16_bn`	`tilman-rassy/doctr-crnn-vgg16-bn-fascan-v1`	French + German + §

Dummy models, handy for tests: Felix92/doctr-dummy-torch-resnet18 (classification), Felix92/doctr-dummy-torch-lw-detr-s (layout) and Felix92/doctr-dummy-torch-tablecenternet (table structure).