Skip to content

tk_r_em 2.0.0 — ONNX Runtime rewrite

Latest
Latest

Choose a tag to compare

View all tags
@Ivanlh20 Ivanlh20 released this 13 Apr 20:57
v2.0.0
380a1f3

tk_r_em 2.0.0

Complete rewrite of the inference engine from TensorFlow/Keras to ONNX Runtime. The six pre-trained CGRDN models are now shipped as .onnx files and run on CPU, NVIDIA GPU, or DirectML without any TensorFlow dependency.

What's new

  • ONNX Runtime backend — no TensorFlow/Keras dependency at runtime. Three install variants: [cpu], [gpu] (NVIDIA CUDA), [directml] (Windows DX12).
  • predict_patch_based — seamless tiled inference for images larger than GPU memory, blended with a Butterworth window.
  • Streamlit web app (app.py) — drag-and-drop browser interface for single-image restoration.
  • EMD file support — Velox/Thermo Fisher .emd files can now be loaded in both the Python API and the Streamlit web app via rsciio.emd.
  • Four example scripts — simulated data, experimental single/batch/patch-based inference.
  • Two Jupyter notebooks — introduction notebook and a full workshop tutorial with inline outputs.
  • pyproject.toml with [cpu], [gpu], [directml], and [notebook] extras.
  • Licence: GPL-3.0-only — all SPDX headers, pyproject.toml, and version.py consistently declare GPL-3.0-only.
  • Fail-loud error handling — removed silent exception swallowing in ONNX DLL preloading and model discovery; invalid user input now shows warnings instead of silently falling back.
  • Standardised file headers — all .py files have the copyright/SPDX/module-docstring block.

Installation

git clone https://github.com/Ivanlh20/tk_r_em.git
cd tk_r_em
pip install -e ".[cpu]"           # CPU (portable)
# pip install -e ".[gpu]"        # NVIDIA CUDA
# pip install -e ".[directml]"   # Windows DX12

Models

Six pre-trained CGRDN networks (7.04 M parameters each), one per modality x resolution:

Tag Modality Resolution
sfr_hrsem SEM High
sfr_lrsem SEM Low
sfr_hrstem STEM High
sfr_lrstem STEM Low
sfr_hrtem TEM High
sfr_lrtem TEM Low

Reference

Lobato, I., Friedrich, T., Van Aert, S. Deep convolutional neural networks to restore single-shot electron microscopy images. npj Computational Materials 10, 10 (2024). doi:10.1038/s41524-023-01188-0

v1 (TensorFlow/Keras) history is preserved via the v1.0.5 tag.

Assets 2
Loading