app.py

import sys
import torch
import onnx
import onnxruntime as rt
from torchvision import transforms as T
from tokenizer_base import Tokenizer
import pathlib
from PIL import Image
from huggingface_hub import Repository


class DocumentParserModel:
    def __init__(
        self,
        repo_path,
        model_subpath,
        img_size,
        charset,
        repo_url="stevenchang/captcha",
        token=None,
    ):
        self.repo_path = pathlib.Path(repo_path).resolve()
        self.model_path = self.repo_path / model_subpath
        self.charset = charset
        self.tokenizer_base = Tokenizer(self.charset)
        self.initialize_repository(repo_url, token)
        self.transform = self.create_transform_pipeline(img_size)
        self.ort_session = self.initialize_onnx_model(str(self.model_path))

    def initialize_repository(self, repo_url, token):
        if not self.model_path.exists():
            if not self.repo_path.exists():
                print(
                    f"Repository does not exist. Cloning from {repo_url} into {self.repo_path}"
                )
                repo = Repository(
                    local_dir=str(self.repo_path),
                    clone_from=repo_url,
                    use_auth_token=token if token else True,
                )
            else:
                print(
                    f"Model does not exist, but repository is already cloned. Pulling latest changes in {self.repo_path}"
                )
                repo = Repository(
                    local_dir=str(self.repo_path),
                    use_auth_token=token if token else True,
                )
                repo.git_pull()
        else:
            print(
                f"Model {self.model_path} already exists, skipping repository update."
            )

    def create_transform_pipeline(self, img_size):
        transforms = [
            T.Resize(img_size, T.InterpolationMode.BICUBIC),
            T.ToTensor(),
            T.Normalize(0.5, 0.5),
        ]
        return T.Compose(transforms)

    def initialize_onnx_model(self, model_path):
        onnx_model = onnx.load(model_path)
        onnx.checker.check_model(onnx_model)
        return rt.InferenceSession(model_path)

    def predict_text(self, image_path):
        try:
            with Image.open(image_path) as img_org:
                x = self.transform(img_org.convert("RGB")).unsqueeze(0)
                ort_inputs = {self.ort_session.get_inputs()[0].name: x.cpu().numpy()}
                logits = self.ort_session.run(None, ort_inputs)[0]
                probs = torch.tensor(logits).softmax(-1)
                preds, _ = self.tokenizer_base.decode(probs)
                return preds[0]
        except IOError:
            print(f"Error: Cannot open image {image_path}")
            return None


if __name__ == "__main__":
    import sys

    repo_path = "secret_models"
    model_subpath = "captcha.onnx"
    img_size = (32, 128)
    charset = r"0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~"

    doc_parser = DocumentParserModel(
        repo_path=repo_path,
        model_subpath=model_subpath,
        img_size=img_size,
        charset=charset,
    )
    if len(sys.argv) > 1:
        image_path = sys.argv[1]
        result = doc_parser.predict_text(image_path)
        print(result)
    else:
        print("Please provide an image path.")