Closed Sukh-P closed 4 weeks ago
Good to note its not a wandb
issue as this works
import wandb
wandb_logger = dict(_target_="lightning.pytorch.loggers.wandb.WandbLogger",
project="wandb_test",
name="test_mode",
save_dir="wandb_test",
offline=False, # set True to store all logs only locally
log_model=False,
job_type="train",
optimizer = dict(_target_="torch.optim.Adam", lr=0.1)
)
experiment = wandb.init(config=wandb_logger)
experiment.log({"loss": 0.1})
It could be a pytorch lightning issue, or a hydra issue
Also doesnt seem to be a problem with pytorch lightning
logger = WandbLogger(
project="wandb_test",
name="test_mode",
save_dir="wandb_test",
offline=False, # set True to store all logs only locally
log_model=False,
job_type="train",
)
logger.log_hyperparams({"optimmizer":{"_target_":"torch.optim.Adam","lr": 0.1}})
logger.experiment.log({ "loss": 0.1})
works, so my guess is its a hydra issue
It looks like its happens in
OmegaConf.save(config.model, f"{callback.dirpath}/model_config.yaml")
in pvnet/training.py
Edit: I'm not sure it in here now
The file seems to appear once the trainer.fit
phase has begun
Whats also weird, is this config seems to save just the just the model
configs, but not other configs
Ill close this now, as the PR #217 should fix this
Describe the bug
Currently when using W&B as the logger when training a model using PVNet not all config parameters seem to show accurately in the W&B UI, some show as
null
even when they are non null.See example screenshot:
To Reproduce
Train a model using PVNet and use W&B as the logger and look at the config pane on the overview page of the run and compare the config that was actually used (should be saved in a local file when running PVNet).
Expected behaviour
Would expect W&B to log all the model config accurately so that comparisons can be made between different runs with different configs (not that currently a local config parameter file is produced so this information is not lost but would be favourable
Additional context
This seems to only happen for section of the config which have formats such as this in the multimodal.yaml file, where
_target_
is the first key on the second level. Most likely is an issue with W&B parsing config in this specific format