Now that we know how to run SLM Lab, let's see how to interpret the data it produces.
As a run (Session, Trial, or Experiment) completes, its data is saved to the data/ folder with an identifying spec name appended with a timestamp, e.g. ppo_lunar_2019_11_30_002958. The folder is also automatically zipped for convenient file transfer. A lab data folder is organized as follows:
graph/ # all the plotted graphs for Session, Trial, Experiment
info/ # all the data: metrics and evaluation dataframes, etc.
log/ # log files
model/ # PyTorch model weights
*graphs # important graphs from graph/ for easy access
*metrics # important metrics from info/ for easy access
*specs # important specs from info/ for easy access
The data files are generated mostly by the lab's analysis module, with self-explanatory names. For most purposes, we are interested in the graphs and metrics files saved directly under the data folder produced by a run. Next, we will look at how to make use of these data files to interpret the lab results corresponding to the Lab Organization.