π Awesome OCR multiple programing languages toolkits based on ONNX Runtime, OpenVINO, MNN, PaddlePaddle, TensorRT and PyTorch.
# -*- encoding: utf-8 -*-
# @Author: SWHL
# @Contact: [email protected]
from pathlib import Path
import torch
from omegaconf import OmegaConf
from ....utils.download_file import DownloadFile, DownloadFileInput
from ....utils.log import logger
from ...base import FileInfo, InferSession
from .architectures.base_model import BaseModel
class ModelLoader:
root_dir = Path(__file__).resolve().parent
DEFAULT_CFG_PATH = root_dir / "arch_config.yaml"
def __init__(self, cfg, device: torch.device):
model_path = self._init_model_path(cfg)
arch_config = self._load_arch_config(model_path)
self.predictor = self._build_and_load_model(arch_config, model_path)
self.predictor.to(device)
self.predictor.eval()
def _init_model_path(self, cfg) -> Path:
model_path = cfg.get("model_path", None)
if model_path is None:
model_info = InferSession.get_model_url(
FileInfo(
engine_type=cfg.engine_type,
ocr_version=cfg.ocr_version,
task_type=cfg.task_type,
lang_type=cfg.lang_type,
model_type=cfg.model_type,
)
)
default_model_url = model_info["model_dir"]
model_path = Path(cfg.model_root_dir) / Path(default_model_url).name
DownloadFile.run(
DownloadFileInput(
file_url=default_model_url,
sha256=model_info["SHA256"],
save_path=model_path,
logger=logger,
)
)
logger.info(f"Using {model_path}")
InferSession._verify_model(model_path)
return Path(model_path)
def _load_arch_config(self, model_path: Path):
all_arch_config = OmegaConf.load(self.DEFAULT_CFG_PATH)
file_name = model_path.stem
if file_name not in all_arch_config:
raise ValueError(f"architecture {file_name} is not in arch_config.yaml")
return all_arch_config.get(file_name)
def _build_and_load_model(self, arch_config, model_path: Path):
model = BaseModel(arch_config)
try:
state_dict = torch.load(model_path, map_location="cpu", weights_only=True)
except TypeError as e:
raise RuntimeError(
"Safe model loading requires PyTorch >= 1.13. Please upgrade PyTorch to load this model securely."
) from e
model.load_state_dict(state_dict)
return model