quantem.widget

quantem.widget#

Interactive, GPU-aware Python widgets for electron microscopy. Use them in Jupyter notebooks, as local HTML files, or from the command line.

Show4DSTEM WebGPU demo with a diffraction pattern and live virtual detector image

Demo: Show4DSTEM HTML with WebGPU. Explore live diffraction-pattern and virtual-detector views locally in a browser on a personal laptop or supported phone, without a Python kernel or remote compute server. Thanks to Serin Lee for sharing this liquid-cell Au nanoparticle 4D-STEM dataset. Check Serin’s 4D-STEM and 5D-STEM segmentation and clustering work (paper) and the source data (Zenodo).

Start with ARINA 4D-STEM in Jupyter#

The demo above is the same Show4DSTEM workflow you can use at the microscope. After installing, open a Jupyter notebook, load a completed ARINA *_master.h5 file, and pass the result directly to the widget:

from quantem.gpu.io import load
from quantem.widget import Show4DSTEM

data = load("/data/session/scan_000_master.h5")
viewer = Show4DSTEM(data)
viewer

load(...) selects CUDA or Apple Metal automatically. Leave viewer as the final line, then move through scan positions or drag the detector to update the virtual image. Continue with the Show4DSTEM tutorial or Load and I/O.

Prefer the command line?#

Point the quantem command at a file or folder when you want the same viewers without writing a notebook:

quantem show2d image.tif         # an image            -> Show2D
quantem show3d ./frames/         # a folder of frames  -> Show3D scrub
quantem show4dstem ./masters/    # 4D-STEM master(s)   -> live viewer (or --html)

It saves to ~/Downloads, opens automatically, and picks the GPU for you. Full details are on the command line page.

Built for two platforms#

We serve two audiences first:

  • macOS on Apple M-chips - the Metal (MPS) GPU.

  • Linux with NVIDIA CUDA - workstations and HPC.

CUDA and MPS are the primary backends. Work stays on the GPU as PyTorch tensors; we avoid NumPy on the hot path. Automatic scientific loading and compute never silently fall back to CPU: an unsupported machine fails with a corrective error. The explicit CPU reference exists for parity tests, while the viewers can still display ordinary NumPy arrays supplied by a user. For large datasets, bin the detector at load (det_bin) to cut memory and speed first paint - see Load and I/O.

Widgets#

Widget

Use it for

Tutorial · API

Show1D

Interactive traces, live reconstruction metrics, line profiles, and linked image snapshots

API

Show2D

One or many 2D images: contrast, FFT, ROIs, line profiles, scale bars

tutorial · API

Show3D

A 3D volume scrubbed slice-by-slice (e.g. a ptychographic object)

tutorial · API

Show3DSlices

Side-by-side slices of a 3D volume across an axis

tutorial · API

Show4DSTEM

4D-STEM: live virtual detectors, multi-master review, and WebGPU HTML export

tutorial · export · API

ShowPtycho

Ptychography aberration review: phase, FFT, BF-count tradeoffs, and WebGPU folder export

API

ShowDiffraction

2D/3D diffraction d-spacing: Bragg spots, rings, center finding, k calibration

tutorial · API

ChooseLattice

Pick an origin and two lattice vectors on a 2D image

tutorial · API

ShowFolder

Folder-level microscopy browser: navigate a session, review thumbnails, select files/folders, and save curation state

tutorial · API

The Tutorials walk through each widget on real public data where practical, with compact synthetic data only where it keeps an example portable. Real tutorial datasets are downloaded from public data hosting such as Hugging Face and cached locally; they are not committed to this repository or bundled into the Python wheel. That keeps clone size and microscope-PC installs small while still letting the rendered docs use realistic microscopy examples. The Show4DSTEM export recipes show how to choose between compact report HTML, interactive raw-4D WebGPU HTML, and terminal exports. The ShowFolder tutorial covers folder browsing workflows and how to save and share widget exports. The API reference documents every parameter, method, and interactive control (and doubles as a UI-test spec for automated agents). All example data here is synthetic or pulled from a public Hugging Face dataset - no private data ships in the docs.

Every widget accepts a NumPy array, a PyTorch tensor (CPU or GPU), or a quantem Dataset (Dataset2d / Dataset3d / Dataset4dstem), pulling calibration and units automatically from the dataset when present.

Offline by default in these docs#

The Show2D, Show3D, and Show3DSlices examples on this site were exported with encoding="uint8", which bakes the display data into the widget as a uint8 stack (4x smaller than float32, and the colormap clamps to 256 levels anyway so it looks identical). The canvas below each example stays fully interactive in this static page with no running kernel: scrub, zoom, change contrast, toggle the FFT - all in the browser. Show4DSTEM goes further: for a small dataset its virtual-detector math runs in WebGPU, so dragging the aperture recomputes the virtual image in the browser. Show4DSTEM exports make the dtype explicit: uint8 is a compact browse payload, while uint16 keeps the wider detector-count range at a larger size. See Show4DSTEM export recipes for when to choose each.

ShowEDS uses the same saved-widget model for synthetic and small cubes in single mode with exact data. For large native EDS/EELS files, the notebook keeps the interactive state while the exact count data stays in a data folder. Portable HTML demos can be exported with count-preserving sum downsampling when full-resolution data would be too large for public sharing.

See Installation to get started.

Citing quantem.widget#

If the quantEM interactive framework—including quantem.widget, GPU-accelerated I/O, analysis, or reconstruction workflows on MPS or CUDA—contributed to your research, please consider citing Lee et al., Interactive Framework for Real-Time 4DSTEM Analysis and Reconstruction, Microscopy and Microanalysis 32 (Supplement 1), ozag053.941 (2026), https://doi.org/10.1093/mam/ozag053.941.

Getting help#