[REQUEST] More control over logging options
jdesman1 opened this issue · comments
Is your feature request related to a problem? Please describe.
D3rlpy has so far been a pleasure to work with. However, I've noticed that for models that take many epochs to train, I sometimes sit on an enormous storage footprint after saving many model checkpoints, which can be inconvenient for myself or others that share the same resource allocation.
Describe the solution you'd like
Perhaps this exists and I missed it in the documentation, but it might be nice to have a parameter that represents "log every n epoch". Being able to use this in the logging adapter classes would be a convenient feature for long training runs, or hyperparameter searches that can take a long time.
@jdesman1 Hi, thank you for the issue. Yes, d3rlpy already provides an option to control this, which is called save_interval
.
https://d3rlpy.readthedocs.io/en/v2.3.0/references/algos.html#d3rlpy.algos.QLearningAlgoBase.fit
If you set save_interval=10
, d3rlpy will save models every 10 epochs.