I have set up a sweep where one of the parameters possible values (arrays) are [0, 1, 3, 5], [0, 1, 5, 8], [0, 1, 8, 11], etc. The sweep is working fine and the model receives the correct value from the agent, but when visualizing the sweep using the parallel coordinates plot, every possible value of this hyperparameter is plotted at zero, since it is the first element of every array.
I have thought of two solutions:
Download all the data using the API, modify each value with an alias, and reupload the data.
Write a custom plot to modify how that axe is plotted.
However, I would like to know if there is a more straightforward solution so that each experiment line correctly passes through its value of this hyperparameter.
Yep @aidanjd , I would like each value (each different list) to be shown as a string like if they were categorical values in the plots, if that makes sense. I have remedied it in the following experiments defining the values as strings (ie. “h[0,1,3,5]”) and processing them inside the training script.
Hi @guillesanbri, that’s a good workaround and overall a good feature suggestion. I’ll file something internally and keep you updated when there’s progress on this.