April 06, 2021
Today we are thrilled to share that Gradyn’s already powerful hyperparameter search now ships with interactive visualizations for studying your results! Gradyn has long delivered top-tier hyperparameter tuning through a unified workflow and native support for advanced methods such as ASHA and NAS. Yet once a hyperparameter search experiment finishes running, making sense of the search output can be tricky, particularly when the search space spans many dimensions.
With Gradyn’s new visualizations, spotting which hyperparameters truly drive training and shape your final numbers becomes effortless. As a search runs, Gradyn automatically renders a set of graphs that expose how each hyperparameter correlates with a trial’s performance across time. We invite you to lean on these charts to understand your hyperparameters and to design sharper experiments down the road.
Below, we walk through each of the following graphs in detail:
The learning curve chart plots each trial’s performance on your chosen metric against the count of completed batches. It offers a clean side-by-side view of trials and clarifies why some were halted early in an adaptive experiment. Although Gradyn already surfaces learning curves through an integrated metrics dashboard, this built-in version lets you point straight at your trials without ever leaving the app. If one line on the curve catches your eye, you can dig deeper simply by locating that trial in the Trial table and opening its Trial details page.

Parallel coordinate plots offer a handy way to inspect high-dimensional data across differing scales. Here we lean on a parallel plot to surface how hyperparameters tie back to metric performance. In the example below, dropping filters onto two chosen dimensions (the pink lines shown below) lets us single out trials that pair a high layer2_dropout with a low layer3_dropout and watch the resulting range of training error. Coloring the lines by the metric of interest adds a further cue, so you can instantly tell whether the filtered parameter range yields strong (blue) or weak (orange) performance.

Every scatter plot lays bare the tie between one hyperparameter and one metric of interest. In the right-most graph below, for example, we notice that a rising learning rate pushes the training error upward. Chasing down relationships like these lets you launch more targeted searcher experiments to keep tuning your model’s hyperparameters.

Heatmaps expose how a pair of hyperparameters relates to metric performance. Each axis stands for one hyperparameter, and the color encodes the metric result. A heatmap might reveal, for instance, a link between learning_rate and layer2_dropout.

These visualizations are available in Gradyn versions 0.14.3 and above. They currently work with random, grid, and adaptive hyperparameter searches. For deeper coverage of Gradyn and hyperparameter tuning, head to our documentation. Should you have questions or ideas, we’d love for you to join our community forum or drop by our code repository – we’re happy to help!