How to set custom training parameter values?

Get more targeted results with IKOSA AI by using custom training parameters.

When training deep learning models, you might need to adjust numerous parameters based on the specific application. We solved this task on your behalf by making IKOSA AI as generic as possible by defining preset parameter values tailored to life science use cases.

In most cases, you don’t have to bother with adjusting parameters. However, sometimes the parameter values predefined in IKOSA do not perfectly align with your research design. In such a case, you need to be able to change parameter values to achieve satisfactory results.

What to expect from this page?

What training parameters can be customized?

The application training has been created for particular life science use cases. When training deep learning models, preset parameter values can be limiting, when it comes to more specific image analysis tasks (for example, various scales or large objects).

IKOSA allows you to override default training parameter values and set custom ones for a more tailored training experience.

Below we provide training parameters whose values can be customized.

Decreasing Image Resolution - Downscale Factor

The parameter downscaleFactor allows you to reduce the resolution of your images for training and prediction using deployed apps.

Fact 1

The downscaleFactor has a positive integer value between 1 and 1,000 (1, 2, 3, …, 1,000). This value determines the resolution at which images are used during training. In other words, the downscaleFactor decreases the resolution of the images before they are fed into the model.

Example: with a downscaleFactor of 2, images will be reduced to half of their original size.

Fact 2

A higher downscaleFactor can result in faster training and prediction when using the deployed IKOSA AI app.

Impact on small objects

Accuracy may decrease when a higher downscaleFactor is applied.

Impact on large objects

If you want to identify large (with a radius greater than 32 pixels) and weakly textured objects, model accuracy can improve when a higher downscaleFactor is applied.

Please note: The optimal choice for the downscaleFactor strongly depends on the use case. This parameter needs to be tuned experimentally. To find its optimal value, you can perform a few QUICK training sessions with different downscaleFactors and check how it impacts your training metrics.

Example: The following image example displays H&E-stained mouse prostates along with annotations for ducts (green) and nuclei (purple).

  • H&E stained mouse prostates with annotations of ducts (green) and nuclei (purple).

When training IKOSA AI to segment the duct structures (green) it is advantageous to apply a downscale factor that reduces the resolution of your images. That is because the ducts are large and the full resolution with fine image details is unnecessary for their segmentation. By doing so, both speed and accuracy can be enhanced.

In the example, we experimentally found that a downscale factor of approximately 5-10 produces good results.

However, when training to segment the nuclei (purple), it is crucial to have the full resolution showing all fine image details. Therefore, a downscale factor of 1 should be used to achieve optimal results.

Fact 3

The preset default downscale factor values used for the different training types are as follows:

  • Semantic Segmentation: 2

  • Instance Segmentation (2D): 1

  • Instance Segmentation (Multichannel): 1

How to set custom values for the training parameters?

Custom values for training parameters can be defined as:

  • Via IKOSA AI Wizard (NO technical knowledge required)

  • Via API (technical knowledge required)

How to set them up via IKOSA AI Wizard?

In the last step of your training session, you can select the downscale factor from the Advanced Settings drop-down section.

How to set them up via API?

When starting an IKOSA AI training via an IKOSA Prisma API call (POST request to /training endpoint, see

Here is what starting training with custom parameter values looks like:

  1. Define your data for the JSON payload of the POST request to the /training endpoint and define custom training parameter values in hyperparamOverrides.
    For example:

    { "data": { ..., "hyperparamPreset": "QUICK", "hyperparamOverrides": { "downscaleFactor": 4 } } }
  2. In the request header, include the appropriate authentication information, such as an API key or token.

  3. Send the POST request to the API endpoint.

  4. The API will then process the request and return a response, including information about the training location or errors.

  5. The created training is shown in your IKOSA account under IKOSA AI > Training Overview.

Now you can start creating your first IKOSA AI application training with custom training parameter values. You will see improvement in your app performance after a couple of experiments!

If you need any help evaluating your app performance, feel free to email us at Copy-paste your training ID into the subject line of your email.

Share this tip! Help others to improve their research outcomes with the help of custom parameter values.

If you still have questions regarding your application training, feel free to send us an email at Copy-paste your training ID into the subject line of your email.

Related articles

Copyright 2016-2021 KML Vision GmbH. IKOSA® is a registered EU trademark.