Page orientation model redownloading when container starts

[DJay921]

Hi, I have deployed a small project on serverless, I have specified the custom models like below in my image. However, I see in my server that whenever my container starts, it is downloading the orientation model and saving to my /root/.cache not sure why. Note that I only see page orientation model downloaded and not the other. What will be the reason for this and how can I avoid this redownloading? Thank :)

from onnxtr.models import detection, recognition, mobilenet_v3_small_page_orientation
from onnxtr.models import ocr_predictor
from onnxtr.models.classification.zoo import page_orientation_predictor
detection_model_path = os.path.abspath('onnx/onnxtr-db-mobilenet-v3-large.onnx')
recognition_model_path = os.path.abspath('onnx/onnxtr-crnn-mobilenet-v3-large.onnx')
page_orientation_model_path = os.path.abspath('onnx/mobilenet_v3_small_page_orientation-60606ce4.onnx')


det_model = detection.db_mobilenet_v3_large(detection_model_path)
rec_model = recognition.crnn_mobilenet_v3_large(recognition_model_path)
custom_page_orientation_model = mobilenet_v3_small_page_orientation(page_orientation_model_path)

ocr = ocr_predictor(det_arch=det_model,
                           reco_arch=rec_model, 
                           assume_straight_pages=True,
                           
                           straighten_pages=True                     )

ocr.page_orientation_predictor = page_orientation_predictor(custom_page_orientation_model)

[felixdittrich92]

Hey 👋;

A quick fix would be:

from onnxtr.models import detection, recognition, mobilenet_v3_small_page_orientation, mobilenet_v3_small_crop_orientation
from onnxtr.models import ocr_predictor
from onnxtr.models.classification.zoo import page_orientation_predictor, crop_orienatation_predictor

detection_model_path = os.path.abspath('onnx/onnxtr-db-mobilenet-v3-large.onnx')
recognition_model_path = os.path.abspath('onnx/onnxtr-crnn-mobilenet-v3-large.onnx')
page_orientation_model_path = os.path.abspath('onnx/mobilenet_v3_small_page_orientation-60606ce4.onnx')
crop_orientation_model_path = ...


det_model = detection.db_mobilenet_v3_large(detection_model_path)
rec_model = recognition.crnn_mobilenet_v3_large(recognition_model_path)
custom_page_orientation_model = mobilenet_v3_small_page_orientation(page_orientation_model_path)
custom_crop_orientation_model = mobilenet_v3_small_crop_orientation(crop_orientation_model_path)

ocr = ocr_predictor(det_arch=det_model,
                           reco_arch=rec_model, 
                           assume_straight_pages=True,
                           straighten_pages=True,
                           disable_crop_orientation=True,  # <-- disable before
                           disable_page_orientation=True,  # <-- disable before
                           )

# Assign custom orientation models
ocr.crop_orientation_predictor = crop_orientation_predictor(custom_crop_orientation_model)
ocr.page_orientation_predictor = page_orientation_predictor(custom_page_orientation_model)

# Reenable orientation detection after loading custom models
ocr._page_orientation_disabled = False
ocr._crop_orientation_disabled = False

[DJay921]

hey @felixdittrich92! many thanks for your response. So using the spaces, I see some parameters that work well for me like this attached screenshot, however I'm a little unsure how these 4 parameters are working under the hood and in what order :)

assume_straight_pages,
                           straighten_pages,
                           disable_crop_orientation
                           disable_page_orientation

[felixT2K]

assume_straight_pages enabled => If your documents contain only straight horizontal text (no rotated or vertical)
straighten_pages enabled will correct rotated documents like in your example the -90 degree rotation
disable_page_orientation enabled will disable the neural network under the hood which is a simple classifier to predict the general orientation of the page/document [0. 90, 180, -90] degrees before it is passed to the fine grained orientation detection (so in short if it is enabled it combines the neural network prediction with the cv algorithm otherwise only the cv algorithm is used)
disable_crop_orientation has only an effect in combination with assume_straight_pages=False (disabled in the space) it classifies the general orientation (again [0, 90, 180, -90]) which is required for the text recognition if you deal with rotated documents without straighten_pages before

Hope this helps :)

[milosacimovic]

I'd like to add that the proposed quick fix didn't work for me as my detection model needed to have assume_straight_pages = False.
so I had to redo additional setting

predictor.assume_straight_pages = assume_straight_pages
predictor.det_predictor.model.assume_straight_pages = assume_straight_pages
predictor.det_predictor.model.postprocessor.assume_straight_pages = assume_straight_pages

[felixdittrich92]

Hi @milosacimovic 👋,

Thanks for the feedback 👍
Also with the latest pypi published version ? Because there was a small bug fixed in v0.6.1 :)