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
Background:
In our current setup, we configure directories for logs, models, checkpoints, and results. They helped a lot with keeping everything ordered. And it is super that we integrated wandb for logging trainings.
For me, wandb is integrated into my workflow to manage the training hyperparameters and output values. Typically, these are stored as JSON files (wandb-[metadata|summary].json) under the directory ${workrootdir}/wandb/${wandb_subdir}, where ./wandb/ is the default value of the keyword in wandb.init(), and ${wandb_subdir} is generated on the fly, e.g. run-20240531_154023-9cce39hr.
Problem Statement:
The wandb logging directories are stored under ${workrootdir}/wandb/${wandb_subdir}. This setup does not allow for easily identifiable directory names and relate them to mace training names, leading to difficulties in accurately locating the correct wandb directories and making personalized postprocess difficult.
Proposed Solution:
I propose to introduce a new command-line argument --wandb_dir that users can specify when running mace/cli/run_train.py. This argument corresponds directly to the dir parameter in the wandb.init() function. By implementing this feature, users will have the flexibility to define custom and more descriptive directory paths for wandb logs, enhancing data organization and accessibility.
The text was updated successfully, but these errors were encountered:
Background:
In our current setup, we configure directories for logs, models, checkpoints, and results. They helped a lot with keeping everything ordered. And it is super that we integrated
wandb
for logging trainings.For me,
wandb
is integrated into my workflow to manage the training hyperparameters and output values. Typically, these are stored as JSON files (wandb-[metadata|summary].json
) under the directory${workrootdir}/wandb/${wandb_subdir}
, where./wandb/
is the default value of the keyword inwandb.init()
, and${wandb_subdir}
is generated on the fly, e.g.run-20240531_154023-9cce39hr
.Problem Statement:
The
wandb
logging directories are stored under${workrootdir}/wandb/${wandb_subdir}
. This setup does not allow for easily identifiable directory names and relate them to mace training names, leading to difficulties in accurately locating the correct wandb directories and making personalized postprocess difficult.Proposed Solution:
I propose to introduce a new command-line argument
--wandb_dir
that users can specify when runningmace/cli/run_train.py
. This argument corresponds directly to thedir
parameter in thewandb.init()
function. By implementing this feature, users will have the flexibility to define custom and more descriptive directory paths forwandb
logs, enhancing data organization and accessibility.The text was updated successfully, but these errors were encountered: