Experiments

An experiment in Edge Impulse is an impulse. A project can contain any number of impulses, and you manage and compare them on the Experiments page. Because every impulse in a project is trained and tested on the same project dataset, differences in the results come from the impulse itself: the window size you chose, the processing block and its parameters, the learning block and its architecture, or the post-processing configuration. To try a different configuration, add an impulse instead of duplicating the project.

​

What is shared and what is not

Each impulse has its own ID and its own name. One impulse is active at a time: the impulse selector in the left navigation bar controls which one, and pages such as impulse design, feature generation, training, live classification, model testing, and deployment all act on the impulse that is currently selected. You can also click an impulse in the experiments table to switch to it. This means you can, for example, run FOMO and MobileNet SSD against the same dataset at the same time: add two impulses, configure one with each learning block, train both, and compare them in the table.

​

Creating an experiment

There are three ways to add an impulse to a project:

• Start from an empty impulse. Use the impulse selector in the left navigation bar to create a new impulse, then add input, processing, and learning blocks as you would for the first impulse in a project.
• Clone an existing impulse. Cloning is the fastest way to change one variable at a time. You can copy just the structure (the blocks and their configuration, with no features or trained models), or clone it completely, which also brings over the generated features and trained model so the new impulse starts in the same state as the original.
• Promote an EON Tuner trial. The EON Tuner searches a parameter space for you. When a trial looks promising, click + Add on that trial to add it to your list of experiments as a full impulse that you can then tune by hand.

A new impulse starts empty: it has no features and no trained model until you generate features and train it. The experiments table shows the state of each impulse, including whether it is fully configured, whether it has been trained, and whether it has become stale because a block parameter changed after the last training run.

​

Comparing experiments

The experiments table lists every impulse in the project as a row, with the metrics and configuration values as columns. Columns are grouped by where the value comes from:

• Impulse metrics: results for the impulse as a whole, such as accuracy and estimated on-device latency and memory.
• Input, processing, learning, and post-processing block configuration: the parameters each block was configured with, so you can see which setting produced which result.
• Learning block metrics: per-block training and validation results.

Use the impulse filter and the column selector in the top right corner of the table to choose which impulses and columns are shown. Click any column header to sort by it, and download the table as CSV or JSON to keep a record of a comparison or to analyze it outside the studio.

​

Experiments and the EON Tuner

You can use experiments and the EON Tuner together. A common workflow is to run the EON Tuner first to narrow down the search space, add the most promising trials as impulses, then refine those by hand and compare them in the experiments table.

​

Working with experiments programmatically

Everything on the Experiments page is available through the Edge Impulse Studio API, which is useful for scripted sweeps and for reporting results into an external experiment tracker. Most impulse endpoints take an optional impulseId parameter; when you omit it, they operate on the project’s default impulse. The same operations are available through the Python API bindings and the JavaScript API bindings.

​

Additional resources

• Impulse design
• EON Tuner
• Model testing
• Increasing model performance