You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
#672 describes replacing several options in the config file with a single option model_path.
The semantics of the model_path option are that it points to a directory generated by vak train. We should formalize this as a file format, ideally baked into the code, e.g. as an attrs class.
Since we don't have existing constraints we can put this in place now to make life easier later--I would rather capture as much info as possible, in a directory, about model provenance, training procedure, etc, instead of needing to put a lot of that into code (like the torchvision weights abstraction).
Currently the "file format" is:
directory
with log file
config file
checkpoints folder
that has "intermittent save" checkpoint as well as "best performance on validation set" checkpoint
parts of transform pipeline, e.g. SpectScaler -- could we save the whole transform pipeline?
for frame classification: labelmap.json
The text was updated successfully, but these errors were encountered:
#672 describes replacing several options in the config file with a single option
model_path
.The semantics of the
model_path
option are that it points to a directory generated byvak train
. We should formalize this as a file format, ideally baked into the code, e.g. as an attrs class.Since we don't have existing constraints we can put this in place now to make life easier later--I would rather capture as much info as possible, in a directory, about model provenance, training procedure, etc, instead of needing to put a lot of that into code (like the torchvision
weights
abstraction).Currently the "file format" is:
The text was updated successfully, but these errors were encountered: