Source code for torchocr.io.pdf

"""PDF loading via PyMuPDF (``import pymupdf``)."""

from pathlib import Path

import pymupdf
import numpy as np
import torch
from torch import Tensor


[docs] def load_pdf(path: str | Path, dpi: int = 200) -> list[Tensor]: """Load a PDF and return one ``(3, H, W) uint8`` RGB tensor per page. Each page is rendered to a PyMuPDF ``Pixmap`` at the requested DPI and converted into a fresh PyTorch tensor that does not alias the pixmap's underlying memory. The pixmap and its sample buffer are released between pages, keeping peak memory bounded by the largest single page rather than the whole document. Args: path: Filesystem path to the PDF. dpi: Render resolution in dots-per-inch. Default 200 -- a sane OCR baseline. PyMuPDF's native unit is the PostScript point (1/72 inch); the conversion is ``Matrix(dpi/72, dpi/72)``. Returns: A list of ``(3, H, W) uint8`` tensors, one per page in document order. Channel order is RGB. """ if dpi <= 0: raise ValueError(f"dpi must be positive; got {dpi}.") matrix = pymupdf.Matrix(dpi / 72.0, dpi / 72.0) pages: list[Tensor] = [] with pymupdf.open(str(path)) as document: for page in document: pixmap = page.get_pixmap(matrix=matrix, colorspace=pymupdf.csRGB, alpha=False) arr_hwc = np.frombuffer(pixmap.samples, dtype=np.uint8).reshape( pixmap.height, pixmap.width, 3 ) arr_chw = np.transpose(arr_hwc, (2, 0, 1)).copy() pages.append(torch.from_numpy(arr_chw)) return pages