Start training

Begin with a prepared training project and reviewed split.

Choose the result

Enter a descriptive Name for the result module. When Type offers a choice, select:

Result

What it contains

Copy module and replace this model

A new copy of the source workflow with the selected model fine-tuned. The other models and scripts are retained.

Create a standalone module with this model

A new module containing only the selected model, without the source workflow’s other steps or measurement scripts.

Models that run within detected structures remain in a copied workflow with their preceding model. Neither choice overwrites the original module. Inference GPU sets the trained model’s default GPU for later analysis; Training GPU controls this training run.

Review the training settings

The form starts with defaults for the selected model. Review these before starting:

Setting

Meaning

Training GPU

The cloud GPU used for this run.

Training batch size

The number of image views processed together. Larger batches use more GPU memory.

Epochs

The number of training cycles. More epochs take more time and compute.

Augmentations

Transformations applied to training images, selected from the offered checkboxes.

Minimum object area

For individual-object models, removes predicted objects smaller than this area during validation and later analysis. Uses square pixels at the model’s resolution; training annotations are unchanged.

Image-view size controls and additional settings depend on the model. Leave their defaults for the first run unless you have a specific experimental reason to change them. Nearby tissue added as context around a small ROI is input only; it does not become labeled training area.

Start and monitor the run

  1. Confirm the result name, split and settings, and check your Modal spending limit.

  2. Select Start cloud training.

  3. Keep QuPath open while HistoForge prepares and uploads the data.

  4. Once the run appears in Jobs, use Open latest details to inspect progress and metrics.

To request cancellation, use Cancel in Jobs when it is offered. Wait for Cancelled before treating the run as stopped; closing QuPath does not cancel it. If submission reports an unconfirmed response, check Jobs before starting another run; the first request may already have been accepted.

Wait for Complete, then review the trained result. A failed or cancelled run is not a completed model.

Export dataset… is optional. It saves the training data and split locally on your computer for use outside the cloud workflow; it does not start training.