WARNING:root:Invalid model URL, using default initialization

[akarshsoman]

when i try to run this code , it shows WARNING:root:Invalid model URL, using default initialization, can u please help me with this
from doctr.models import ocr_predictor
from doctr.io import DocumentFile

model = ocr_predictor(reco_arch='sar_resnet31', pretrained=True)
multi_img_doc = DocumentFile.from_images(["2.jpg", "4.jpg"])
result = model(multi_img_doc)
result.show(multi_img_doc)
result

[felixdittrich92]

@akarshsoman Hi 👋 ,
which framework do you use ? (PyTorch has currently no pretrained for sar_resnet31)

[aditya-ihx]

I still get this error.
I am on v0.7.0.

model = ocr_predictor(det_arch='db_resnet50', reco_arch='parseq', pretrained=True)
# # PDF
doc = DocumentFile.from_images("103291164/page_1.jpeg")
# # Analyze
result = model(doc)

Output is WARNING:root:Invalid model URL, using default initialization.

If this is a problem with pretrained version, can I please know which models do have functional pretrained versions. Could not find such a thing in the docs

[felixdittrich92]

Hi @aditya-ihx

v0.7.0:

TF: all models have pretrained checkpoints

PT:
Detection:
db_resnet50, db_mobilenet_v3_large, db_resnet50_rotation
Recognition:
crnn_vgg16_bn, crnn_mobilenet_v3_small, crnn_mobilenet_v3_large

main branch: (v0.7.1a0 currently)

PT:
Recognition:
master, sar_resnet31, vitstr_small, vitstr_base, parseq

We are still busy with training runs but with the next release each model will have a pretrained checkpoint :)

[aditya-ihx]

Since parseq for PT is available, what would be the reason I get the warning WARNING:root:Invalid model URL, using default initialization.

Thanks in advance

[felixdittrich92]

You mentioned that you are on v0.7.0 atm. but parseq is only available at the main branch :) (v0.7.1)
So for parseq you need to clone the repo and install it locally

[aditya-ihx]

Oh! I am sorry. That was a pretty dumb question.

Thank you so much :)

[mehdi-jamaseb]

I hope this message finds you well. I am currently working on an OCR project using the doctr library and encountered an issue when trying to use the fast_tiny detection architecture. Despite setting pretrained=True, I receive the following warning:

WARNING:root:Invalid model URL, using default initialization.
Here is the relevant part of my code:

from doctr.models import ocr_predictor

det_arch = "fast_tiny"
rec_arch = "crnn_vgg16_bn"
predictor = ocr_predictor(det_arch=det_arch, reco_arch=rec_arch, pretrained=True)

Could you please advise on how I can resolve this issue so that the pretrained model is correctly downloaded and used? I have ensured that I am using the latest version of doctr.

Thank you in advance for your assistance

[mehdi-jamaseb]

Dear @felixdittrich92 ,
Thank you for your response and the information regarding the availability of the fast checkpoints in the upcoming v0.9.0 release.
I am currently using Colab for my project. Could you please provide guidance on how I can switch to the main branch (v0.9.0a0) in Colab to test the fast checkpoints?
Additionally, if I were to manually update the URL for the pretrained model, could you please advise on the steps or the specific URL format I should use?
Thank you once again for your assistance.

[mehdi-jamaseb]

thanks @SkaarFacee

[felixT2K]

@mehdi-jamaseb
In colab you can install from the main branch with

!pip3 install python-doctr[torch]@git+https://github.com/mindee/doctr.git

to load custom models please take a look at:
https://mindee.github.io/doctr/using_doctr/custom_models_training.html#loading-your-custom-trained-model

So you would need first to download the model and then load as described in the docs :)

PS: the fast model has a minor bug in v0.8.1 so i suggest to use the main branch ftm

[mehdi-jamaseb]

@felixT2K
Thank you for the instructions.
Thanks again for your help and the bug notice in v0.8.1

[felixT2K]

If you are looking for something more lightweight you could also take a look at: https://github.com/felixdittrich92/OnnxTR
(the last release contains already all features from the current doctr main branch)