Configuration#

docTR has no configuration file: predictors are configured in code, and a few environment variables control how the library interacts with the system.

Environment variables#

Variable	Default	Effect
`DOCTR_CACHE_DIR`	`~/.cache/doctr`	Where downloaded model weights (`models/`) and datasets (`datasets/`) are stored
`DOCTR_MULTIPROCESSING_DISABLE`	unset	Set to `TRUE` to disable the thread pools used for data processing, e.g. where `/dev/shm` is not writable

Model cache#

Pretrained weights are downloaded the first time a model is used, then loaded from the cache. Their integrity is checked with the hash in the file name.

export DOCTR_CACHE_DIR=/opt/models/doctr

Offline use#

To run on a machine without internet access, populate the cache on a connected machine, then copy it:

# On a connected machine: instantiating the predictors downloads their weights
from doctr.models import ocr_predictor

ocr_predictor(pretrained=True)  # detection + recognition
ocr_predictor(pretrained=True, assume_straight_pages=False)  # + orientation classifiers
ocr_predictor(pretrained=True, detect_tables=True)  # + layout and table models

Copy the cache directory (~/.cache/doctr or $DOCTR_CACHE_DIR) to the same location on the offline machine. Alternatively, load the weight files from any path with model.from_pretrained("path/to/weights.pt") (see Using your trained models).

AWS Lambda#

The AWS Lambda runtime only allows writing to /tmp, and does not provide the shared memory used by multiprocessing. Configure the function with:

DOCTR_MULTIPROCESSING_DISABLE=TRUE
DOCTR_CACHE_DIR=/tmp/doctr

Tip

Downloading the weights at every cold start is slow. Package them in your container image or a Lambda layer, and point DOCTR_CACHE_DIR to their location if it is writable, or load them with from_pretrained.

The same settings apply to other restricted environments (read-only containers, some serverless platforms).

Logging#

docTR logs through the standard logging module, under the doctr logger:

import logging

logging.getLogger("doctr").setLevel(logging.WARNING)