crnn_vgg16_bn model - Link for downloading

[gganes3]

Hi All,

Just like the dbresnet (https://github.com/mindee/doctr/releases/download/v0.2.0/db_resnet50-adcafc63.zip) model file, I am trying to download the crnn_vgg16_bn model to use it offline. Could you please provide me the link?

--
Regards,
Gokul.

[charlesmindee]

Hi @gganes3,
the download link is in the cfg of the model:
here for the pytorch version and
here for the tensorflow version

Have a nice day!

[fg-mindee]

Hello @gganes3 👋

I'm not sure where you found the URL of db_resnet50 but it's exactly the same process for other models!
To add up on Charles' answer, I'd recommend the following if you happen to be looking for other model's checkpoint URLs:

• navigate to the "models" page of the documentation: https://mindee.github.io/doctr/latest/models.html
• find the model you're interested in and click on "SOURCE" which will take you to the source code.
• at the top of the page you'll see the URLs for all pretrained models 👍

If you're familiar with github, no need to go through the documentation, and you will find the crnn URLs:

• for pytorch here: https://github.com/mindee/doctr/blob/main/doctr/models/recognition/crnn/pytorch.py#L29
• for TensorFlow there: https://github.com/mindee/doctr/blob/main/doctr/models/recognition/crnn/tensorflow.py#L28

Take care!

[fg-mindee]

This might be out of my skills. But I just realized something: are you positive that this is a PyTorch checkpoint you're loading here?
Considering the "invalid load key", and that you haven't yet tried to load it on the model, I can only see the explanation that the checkpoint is not correctly formatted for PyTorch. So my guess is that it's either corrupted or that it's TF checkpoint (or something else)

How did you save your checkpoint by the way? the default method from the reference scripts?

[gganes3]

Sorry, it was my mistake. Thanks for giving me the right direction. I was not using the torch model. I downloaded the .pt files and now I can load the .pt files. However, one more question, how should I instantiate the det_model and rec_model in the below code.

   # Instantiate your model
    **det_model= ??
    reco_model = ??**
    det_params = torch.load(r"C:\Projects\db_resnet50-ac60cadc.pt", map_location="cpu")
    reco_params = torch.load(r"C:\Projects\crnn_vgg16_bn-9762b0b0.pt", map_location="cpu")
    # det_model.load_state_dict(det_params)
    # reco_model.load_state_dict(reco_params)

    # Ask the preprocessor of each task to resize and normalize similarly to your training
    # cf. https://github.com/mindee/doctr/blob/main/references/detection/train_pytorch.py#L94 & https://github.com/mindee/doctr/blob/main/references/detection/train_pytorch.py#L109
    det_predictor = DetectionPredictor(
        PreProcessor((1024, 1024), batch_size=1, mean=(0.798, 0.785, 0.772), std=(0.264, 0.2749, 0.287)), det_model)
    # cf. https://github.com/mindee/doctr/blob/main/references/recognition/train_pytorch.py#L97 & https://github.com/mindee/doctr/blob/main/references/recognition/train_pytorch.py#L111
    reco_predictor = RecognitionPredictor(
        PreProcessor((32, 128), preserve_aspect_ratio=True, batch_size=32, mean=(0.694, 0.695, 0.693),
                     std=(0.299, 0.296, 0.301)), reco_model)

    model = OCRPredictor(det_predictor, reco_predictor)
    result = model([image_rgb])

[fg-mindee]

Rather easy then :)

from doctr.models.detection import db_resnet50
from doctr.models.recognition import crnn_vgg16_bn

det_model = db_resnet50(pretrained=False, pretrained_backbone=False)
reco_model = crnn_vgg16_bn(pretrained=False, pretrained_backbone=False)

and assuming you haven't changed the vocab, or any other params of both models

[gganes3]

Thanks, but I get this error 'det_model.load_state_dict([det_params])
AttributeError: 'DBNet' object has no attribute 'load_state_dict''

  # Instantiate your model
    det_model = db_resnet50(pretrained=False)
    reco_model = crnn_vgg16_bn(pretrained=False)
    det_params = torch.load(r"C:\Projects\db_resnet50-ac60cadc.pt", map_location="cpu")
    reco_params = torch.load(r"C:\Projects\crnn_vgg16_bn-9762b0b0.pt", map_location="cpu")
    # det_model.load_state_dict()
    det_model.load_state_dict(det_params)
    reco_model.load_state_dict(reco_params)


    det_predictor = DetectionPredictor(
        PreProcessor((1024, 1024), batch_size=1, mean=(0.798, 0.785, 0.772), std=(0.264, 0.2749, 0.287)), det_model)

    reco_predictor = RecognitionPredictor(
        PreProcessor((32, 128), preserve_aspect_ratio=True, batch_size=32, mean=(0.694, 0.695, 0.693),
                     std=(0.299, 0.296, 0.301)), reco_model)

    model = OCRPredictor(det_predictor, reco_predictor)

[fg-mindee]

I'm not able to reproduce this locally, so my guess is:

• either you don't have the correct version of docTR installed (in this case, update it)
• or you're not running docTR with the correct backend (by default it's TensorFlow). To enable this, add this env variable USE_TORCH at the very top of your script:

import os

os.environ['USE_TORCH'] = '1'

Hopefully that will do it 👍