Local Inference

Local (on-disk) inference is activated by clicking the 📁 button in the main Napari-OmniEM panel. In this mode, data are loaded from local files using dask.array, and intermediate results are currently stored as temporary HDF5 files (this is the current implementation; more efficient backends are planned).

Compared with in-memory inference, local inference:

  • Enables multi-GPU parallel inference when multiple GPUs are available

  • Requires approximately 5–10× the original data size to store temporary files

  • Supports saving outputs in multiple formats, with optional visualization in napari

  • Does not currently support the restoration→segmentation pipeline in a single run (restored images must be saved and reloaded for subsequent segmentation)

Quick Start (Minimal Steps)

  1. Open napari and launch Plugins → OmniEM.
  2. Click 📁 to open the Local Inference panel.
  3. Select input Data, Task, and Solution.
  4. Click Run, confirm the settings, and start inference.
  5. View results in napari or save them to disk.

For advanced configuration, see the detailed steps below.

Step 0: Open Local Inference Panel

  • Click the 📁 icon in the main Napari-OmniEM panel.
  • This option is enabled only when a compatible GPU is available.

Step 1: Select Data

  • Click Browse File to select a local image or volume.
  • Supported formats are those compatible with dask.array.
  • Only grayscale images (no color channels) are accepted.
  • If the data are 3D, specify the z-dimension:
    • By default, the axis with the minimum side length is selected.
    • Override this if your data use a different axis convention.

Step 2: Select Task and Solution

  • Task and solution selection is the same as in in-memory inference.
  • Only solutions compatible with the selected data will be shown.

Step 3: Advanced Settings (Sliding Window & Devices)

In addition to standard sliding window parameters, local inference provides two extra options:

  • Device selection

    • GPU-only inference
    • Multiple GPUs can be selected
    • At least one GPU must be chosen

  • Temporary folder

  • Specify the directory used to store temporary on-disk files during inference

For details on sliding window parameters, see the Sliding Window Parameters page.

Step 4: Run Inference

Before starting inference, the dialog displays: - Estimated number of batches

- **Region of interest (ROI) size**

Steps: - Click Run to start inference.

- When inference finishes, **Run** changes to **Redo**, allowing you to rerun with modified settings.

Step 5: Save Output

Supported output options include:

  1. Open in napari

  2. Both input data and results can be loaded

  3. Use with caution for very large datasets

  4. Standard file formats

  5. TIFF

  6. PNG (2D results only)

  7. NumPy (.npy)

  8. Chunked storage formats

  9. Zarr

  10. NumPyStack

Tips & Common Pitfalls

  • Ensure sufficient disk space for temporary files (5–10× input size).
  • For very large volumes, prefer smaller batch sizes to reduce peak memory usage.
  • Clean up temporary folders before inference if disk space is limited.
  • Temporary files are be removed automatically after inference.