I’ve got a large nested config file that I usually log with each runs. The problem is that I only see the beginning of each of the path of the hyperparameters, which isn’t so useful.
Hi @lthiet, thanks for writing in! Currently this isn’t possible as you need to have one unique identifier for each column so that’s why the full path is showed. I can create a new feature request for being able to change the name of columns in the Runs Table if you want, could you explain me a little bit about your use-case? Thanks!
Sure. To configure my runs, I use a JSON file. It has some levels of indirection. For example at the top level I have two nested JSON objects that describes the configuration of the training pipeline (e.g. epoch, etc.) and the configuration of the model (layer size, learning rate, etc.). This is a trivial example and in reality I must have 3 or 4 nested objects until I get to an actual parameters. For example the path to learning rate in my configuration file is as such: vdbae_config.model_config.unet_model_config.optimizer_scheduler_config.learning rate. Usually, only the last part of the path matters to me (learning rate) so it would be nice to display that instead of the whole path.
I would also like to have this feature in wandb. I had 2 tensorboard runs with their directory names different and so after uploading, when going to the runs table to compare different logged metrics, the column names are different (prefixed by the folder name). So we can’t see the results of the same metric under one column.
Is would be great if it is possible to rename a column to some generic name like accuracy, so that the column can have values for both.