In-memory inference

In-memory inference is performed directly within the main Napari-OmniEM panel and operates on image data loaded as napari image layers. All computation is executed in memory, making this mode ideal for interactive exploration, debugging, and small-to-medium EM datasets that fit into GPU/CPU memory.

Quick Start (Minimal Steps)

  1. Open napari and load an EM image or volume.

  2. Open Plugins → OmniEM.

  3. Click 🔄 Refresh / Register to detect loaded data.

  4. Select Data, Task, and Solution.

  5. Click Run → confirm → view results in napari.

For more control and advanced settings, see the detailed steps below.

Step 1: Select data

  • Load image data into napari via File → Open File(s).

    • The data will appear as one or more image layers.

  • Click the 🔄 Refresh / Register button in the Napari-OmniEM panel to register the currently loaded image layers.

  • Select the target image from the Data dropdown list.

  • If the selected data is a 3D volume, specify the z-dimension:

    • By default, the axis with the minimum side length is automatically chosen.

    • You can override this if your data uses a different axis convention.

Step 2: Select task and solution

  • Choose a Task corresponding to the intended processing goal (e.g., segmentation, restoration).

  • Select a compatible Solution from the Solution list.

    • Only solutions that are compatible with the selected data (dimensionality, modality, and task) will be shown.

    • For detailed descriptions of available tasks and solutions, refer to the Model Zoo.

Terminology note - Task: defines what problem is being solved. - Solution: a concrete model + configuration that solves a task.

This terminology is consistent across in-memory and local inference modes.

Step 3 advanced settings for sliding window inference

  • In-memory inference uses the same sliding window mechanism as local inference.

  • Adjust sliding window hyperparameters (e.g., input size, overlap, batch size) as needed.

  • For detailed explanations and recommended values, see:

If you are unsure, the default parameters are generally safe for first-time use.

Step 4: Run inference

  • Click the Run button to open the inference dialog.

  • Review:

    • Estimated number of batches

    • Region-of-interest (ROI) size

  • Click Run in the dialog to start inference.

  • Once inference finishes, close the dialog to view results directly in napari as new layers.

Tips & Common Pitfalls

Memory usage

  • In-memory inference loads data and intermediate results into memory.

  • For large 3D volumes, GPU memory can be exhausted quickly.

  • If you encounter out-of-memory errors:

    • Reduce batch size (even 1 is acceptable)

    • Reduce input window size

    • Consider using Local Inference instead

Batch size

  • Batch size primarily affects memory usage, not output quality.

  • Small batch sizes are recommended for large EM data.

Refresh/Register button

  • Always click 🔄 Refresh / Register after:

    • Loading new data

    • Removing image layers

    • Renaming layers

  • Failure to do so may result in missing or outdated data selections.

When to use Local Inference instead

Use Local Inference if:

  • The dataset is too large to fit into memory

  • You want multi-GPU parallel inference

  • You want to save results directly to disk (e.g., TIFF, Zarr, Dask)