Enable use of LazyConfigs with the existing Defaultrainer
Motivation & Examples
DefaultTrainer is quite nice. Straight from the documentation "it includes more standard default behaviors that one might want to opt in, including default configurations for optimizer, learning rate schedule, logging, evaluation, checkpointing etc.". However, it is limited to yacs configs.
If one is using the new lazyconfigs, the advice is to copy paste the code in lazyconfig_train_net.py and start hacking at it, which can be a bit daunting. I think it would make things simpler if one could just use DefaultTrainer to begin with and then expand as needed. Effectively the same approach one would take with YACS configs.
Suppose that DefaultTrainer did not exist. The advice for someone with a YACS config would then be to copy a tools/train_net.py (a much more complicated one - the current one effectively calls DefaultTrainer) and start from there. Having DefaulTrainer simplifies this. Unfortunately, this option does not exist if one is starting with lazyConfig files.
Having had a look at it, I feel that it should be possible (a lot of the code in DefaultTrainer is similar to the code in lazyconfig_train_net.py). There's a few cases that I'm unsure about but if I know that there is interest by upstream, then I could spend time working on it.
Please let me know if there's interest on extending DefaultTrainer to support LazyConfig.
hiyyg
commented
6 months ago
š Feature
Enable use of LazyConfigs with the existing
DefaultrainerMotivation & Examples
DefaultTraineris quite nice. Straight from the documentation "it includes more standard default behaviors that one might want to opt in, including default configurations for optimizer, learning rate schedule, logging, evaluation, checkpointing etc.". However, it is limited to yacs configs.If one is using the new lazyconfigs, the advice is to copy paste the code in
lazyconfig_train_net.pyand start hacking at it, which can be a bit daunting. I think it would make things simpler if one could just useDefaultTrainerto begin with and then expand as needed. Effectively the same approach one would take with YACS configs.Suppose that
DefaultTrainerdid not exist. The advice for someone with a YACS config would then be to copy atools/train_net.py(a much more complicated one - the current one effectively callsDefaultTrainer) and start from there. HavingDefaulTrainersimplifies this. Unfortunately, this option does not exist if one is starting with lazyConfig files.Having had a look at it, I feel that it should be possible (a lot of the code in
DefaultTraineris similar to the code inlazyconfig_train_net.py). There's a few cases that I'm unsure about but if I know that there is interest by upstream, then I could spend time working on it.Please let me know if there's interest on extending
DefaultTrainerto support LazyConfig.