Miscellaneous apps for daily problems

html2md

convert html input/input.html to markdown output/output.md

decode-encode

decode or encode a string

input for encode: input/encode.txt

output for encode: output/encoded.txt

input for decode: input/decode.txt

output for decode: output/decoded.txt

pdf-ocr

requirements

sudo dnf install tesseract

html2text

convert html input/input.html to markdown output/html2text-result.txt

pdf2md

Convert PDF documents to Markdown. Keeps headings, paragraphs, lists, bold and italic text, links, code blocks and simple tables, reads multi-column pages in column order, and drops repeated page headers and footers. Scanned pages go through OCR.

sudo dnf install mupdf tesseract tesseract-langpack-ind   # mupdf recommended, tesseract for OCR

make install                      # build and install into ~/.local/bin

pdf2md -i report.pdf                          # Markdown to stdout
pdf2md -i report.pdf -o report.md
pdf2md -i pdfs/ -o mds/                       # whole directory
pdf2md -i book.pdf -o book.md -pages 1-5,9 -images
pdf2md -i scan.pdf -o scan.md -ocr always -lang ind+eng

~/.local/bin must be on PATH (it is by default on Fedora). Use make install BINDIR=/some/dir to install elsewhere, and make uninstall/pdf2md to remove it.

Text is extracted with MuPDF's mutool. Without it, a pure Go backend is used (-backend native), which handles fewer PDFs. Run pdf2md -h for all flags.

rmbg

Remove the background from images with a local ONNX segmentation model. Runs on CPU, and on NVIDIA GPUs when the CUDA ONNX Runtime build is used.

make ort/cpu                      # download ONNX Runtime into ./lib (CPU)
make build/rmbg                   # build ./bin/rmbg

./bin/rmbg -i input/photo.jpg -o output/photo.png
./bin/rmbg -i input/photos/ -o output/photos/     # whole directory
./bin/rmbg -i photo.jpg -o mask.png -mask-only    # alpha mask only
./bin/rmbg -i photo.jpg -o out.jpg -bg "#ffffff"  # solid background

The first run downloads the model into $XDG_CACHE_HOME/rmbg (default ~/.cache/rmbg) and verifies its SHA-256.

Models: u2net (default), u2netp, u2net_human_seg, isnet-general-use. Pass -model /path/to/model.onnx to use a custom model.

CPU install (any x86_64 Linux)

make ort/cpu

This unpacks libonnxruntime.so into ./lib, next to the binary. If you install it elsewhere, point the tool at it with -ort-lib or the ONNXRUNTIME_LIB environment variable.

NVIDIA install (Fedora 44)

ONNX Runtime 1.31 GPU builds target CUDA 13.x (CUDA 13.0 or newer, so the CUDA 13.4 toolkit works) and require cuDNN 9.

1. Install the NVIDIA driver, then the CUDA 13.4 toolkit from the Fedora 44 CUDA repo:

   sudo dnf install -y dnf-plugins-core
   sudo dnf config-manager addrepo \
     --from-repofile=https://developer.download.nvidia.com/compute/cuda/repos/fedora44/x86_64/cuda-fedora44.repo
   sudo dnf install -y cuda-toolkit-13-4

   Add CUDA to your environment:

   export PATH=/usr/local/cuda/bin:$PATH
   export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

2. Install cuDNN 9 for CUDA 13 (not shipped in the CUDA repo). Download the archive and put its libraries on LD_LIBRARY_PATH:

   curl -fLO https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-9.27.0.42_cuda13-archive.tar.xz
   sudo tar -xJf cudnn-linux-x86_64-9.27.0.42_cuda13-archive.tar.xz -C /usr/local
   export LD_LIBRARY_PATH=/usr/local/cudnn-linux-x86_64-9.27.0.42_cuda13/lib:$LD_LIBRARY_PATH

3. Download the CUDA 13 build of ONNX Runtime:

   make ort/gpu              # ORT_CUDA=cuda13 by default; ORT_CUDA=cuda12 for CUDA 12.8

4. Run with -device cuda (or -device auto). The CUDA and cuDNN libraries must be on LD_LIBRARY_PATH, and ONNX Runtime's libonnxruntime_providers_* files sit next to libonnxruntime.so in ./lib.

The CUDA, cuDNN, and ONNX Runtime versions must match the release pinned in the Makefile (ORT_VERSION, ORT_CUDA). -device auto falls back to CPU with a warning when CUDA is unavailable; -device cuda fails with exit code 2 instead.

If CUDA is selected but inference fails with CUDNN_STATUS_SUBLIBRARY_VERSION_MISMATCH, more than one cuDNN 9 release is visible and they are being mixed. Keep the loader and all libcudnn_*.so.9 sublibraries from a single cuDNN release, with that directory before any other directory that ships its own cuDNN (for example an ArrayFire install): LD_LIBRARY_PATH=/opt/cudnn/current/lib:/usr/local/cuda/lib64. make run/rmbg prepends CUDNN_LIB_DIRS and CUDA_LIB_DIRS (defaults /opt/cudnn/current/lib and /usr/local/cuda/lib64) for this reason; override them to point at your install.