unknown architecture: Integrate trained model in streamlit app.py

[khawar-islam]

I have trained a small master model and i want to integrate in your streamlit demo/app.py file. I tried to pass the dictionary of my trained model but i dont know where i have to give the trained model path.
Traceback

ValueError: unknown architecture '/media/cvpr/CM_2/doctr/master_20220916-173609.pt'
Traceback:
File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/streamlit/runtime/scriptrunner/script_runner.py", line 556, in _run_script
    exec(code, module.__dict__)
File "/media/cvpr/CM_2/doctr/demo/app.py", line 116, in <module>
    main(DET_ARCHS, RECO_ARCHS)
File "/media/cvpr/CM_2/doctr/demo/app.py", line 84, in main
    predictor = load_predictor(det_arch, reco_arch, forward_device)
File "/media/cvpr/CM_2/doctr/demo/backend/pytorch.py", line 21, in load_predictor
    predictor = ocr_predictor(
File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/doctr/models/zoo.py", line 97, in ocr_predictor
    return _predictor(
File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/doctr/models/zoo.py", line 40, in _predictor
    reco_predictor = recognition_predictor(reco_arch, pretrained=pretrained, batch_size=reco_bs)
File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/doctr/models/recognition/zoo.py", line 59, in recognition_predictor
    return _predictor(arch, pretrained, **kwargs)
File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/doctr/models/recognition/zoo.py", line 24, in _predictor
    raise ValueError(f"unknown architecture '{arch}'")

[felixdittrich92]

Hi @khawar-islam i have posted in #1029 last comment how you can load your custom trained model

[khawar-islam]

@felixdittrich92 Yes, I have checked that comment. I would like to integrate trained model into streamlit app. Is that possible?

[felixdittrich92]

example to load custom model in PT:

import torch

from doctr.datasets import VOCABS
from doctr.io import DocumentFile
from doctr.models import ocr_predictor, master

doc = DocumentFile.from_images("/home/felix/Desktop/1.jpg")
model = master(pretrained=False, vocab=VOCABS["korean"])  <--- your created VOCAB
model.load_state_dict(torch.load("/media/cvpr/CM_2/doctr/master_20220916-173609.pt", map_location="cpu"))
predictor = ocr_predictor(reco_arch=model, det_arch="db_mobilenet_v3_large", pretrained=True)

res = predictor(doc)
print(res)

If GPU is available change this line:
predictor = ocr_predictor(reco_arch=model, det_arch="db_mobilenet_v3_large", pretrained=True).cuda()

Sure you can build your own streamlit app :) With our demo this will not work in fact that matplotlib can't display the korean symbols/chars