I have developed a script to estimate some hardware evaluations for running my neural network. As parameter for this script the number of layers can be defined. This parameter is also passed to Wand.config.
Then the network trains over several epochs and evaluates the hardware. The results are all logged. At the end I would like to have a line chart with the number of layers as xaxis and for example the energy consumption on the yaxis. However, it is not possible to select the num_layers parameter in the menu.
My guess is that line charts can only be plotted across a run and not as in my case where per value of the xaxis (num_layers) a different run specifies the value.
Is there any way to implement my plan in Wandb in an automated way? Otherwise I would have to export the data and plot it with matplotlib etc.
Hi @mohammadbakir! You can have a look here: Weights & Biases
In this workspace there are 6 runs; 3 for each learning rule which I want to compare. These two sets of runs have an increasing network depth (2, 3, 4). Now I’d like to plot different metrics such as accuracy over the different network depts. So, on the x-axis there would be the depth with values 2, 3, 4; while measuring accuracy on the y-axis.
All this should be plotted as a line chart with one line for each learning rule, i.e. comparing “BP” with “DFA”.
Hi @chwolters , this is definitely doable with custom charts. Specifically a chart with a custom x axis of 2,3,4, and three y fields values to referencing the runs accuracy. Alternatively yes, using Matplotlib is an option, however, with custom charts, you can save the preset for future use.
If you decide for a custom chart, some of your Vega code may appear as follows
Hi @chwolters , since we have not heard back from you we are going to close this request. If you would like to re-open the conversation, please let us know!