captain-pool / hydra-wandb-sweeper

WandB sweeps integration with Hydra sweeper
47 stars 6 forks source link

WandB sweeper integration with Hydra

Installing

$ pip install git+https://github.com/captain-pool/hydra-wandb-sweeper.git

Usage

Using this plugin is very simple, as one can just invoke the sweeper by overriding hydra/sweeper with wandb.

@hydra.main(config_path=..., config_name=...)
def main(cfg: DictConfig):

  with wandb.init(id=run_id) as run:
    ...
    wandb.log({"loss": loss})

Hydra Overrides

Passing Distributions for continuous parameters

Sweeping over continuous params (for ex. learning rates in ML pipelines) can be done by passing the upper limit and lower with the hydra's interval(<min>, <max>) override. For example:

$ python3 /path/to/trainer/file hydra/sweeper=wandb model.learning_rate="interval(0.1, 0.2)"

By default, the interval() sweep uses the uniform distribution while making parameter suggestions. To use a different distribution add the name of the supported distribution (uniform, log_uniform, q_uniform, q_log_uniform, q_normal, log_normal, q_log_normal) as a tag to the interval() sweep. For example, to use a normal distribution with mean=0 and variance=1 to suggest a value for learning_rate use the following:

$ python3 /path/to/trainer/file hydra/sweeper=wandb model.learning_rate="tag(normal, interval(0, 1))"

Categorial parameters

Sweeping over categorical items is as simple as passing a list of items to sweep over thereby directly using Hydra' choice() sweep feature. for example, to sweep over a categorical list of batch_size,

$ python3 /path/to/trainer/file hydra/sweeper=wandb model.batch_size=8,10,12,14

Equivalently, the choice() command can also be expliclty used:

$python3 /path/to/trainer/file hydra/sweep=wandb model.batch_size="choice(8,10,12,14)".

Categorical sweep can also be done using hydra's range() sweep. The previous task can be equivalently achieved as:

$ python3 /path/to/trainer/file hydra/sweeper=wandb model.batch_size="range(8, 15, step=2)"