OmniSafe Logger#

Logger(output_dir, exp_name[, output_fname, ...])

Implementation of the Logger.

Logger#

Documentation

class omnisafe.common.logger.Logger(output_dir, exp_name, output_fname='progress.csv', seed=0, use_tensorboard=True, use_wandb=False, config=None, models=None)[source]#

Implementation of the Logger.

A logger to record the training process. It can record the training process to a file and print it to the console. It can also record the training process to tensorboard.

The logger can record the following data:

----------------------------------------------
|       Name      |            Value         |
----------------------------------------------
|    Train/Epoch  |             25           |
|  Metrics/EpCost |            24.56         |
|  Metrics/EpLen  |            1000          |
|  Metrics/EpRet  |            13.24         |
|  Metrics/EpStd  |            0.12          |
----------------------------------------------
Parameters:

  • output_dir (str) – The output directory.

  • exp_name (str) – The experiment name.

  • output_fname (str, optional) – The output file name. Defaults to ‘progress.csv’.

  • seed (int, optional) – The random seed. Defaults to 0.

  • use_tensorboard (bool, optional) – Whether to use tensorboard. Defaults to True.

  • use_wandb (bool, optional) – Whether to use wandb. Defaults to False.

  • config (Config or None, optional) – The config. Defaults to None.

  • models (list[torch.nn.Module] or None, optional) – The models. Defaults to None.

Initialize an instance of Logger.

_update_current_row()[source]#

Update the current row.

Update the current row with the data stored in the logger.

Return type:

None

close()[source]#

Close the logger.

Return type:

None

property current_epoch: int#

Return the current epoch.

dump_tabular()[source]#

Dump the tabular data to the console and the file.

The dumped data will be separated by the following steps: :rtype: None

  • If the key is registered with window_length, the data will be averaged in the window.

  • Write the data to the csv file.

  • Write the data to the tensorboard.

  • Update the progress logger.

get_stats(key, min_and_max=False)[source]#

Get the statistics of the key.

Parameters:

  • key (str) – The key to be registered.

  • min_and_max (bool, optional) – Whether to record the min and max value of the key. Defaults to False.

Returns:

The mean value of the key or (mean, min, max, std)

Return type:

tuple[float, ...]

log(msg, color='green', bold=False)[source]#

Log the message to the console and the file.

Parameters:

  • msg (str) – The message to be logged.

  • color (str, optional) – The color of the message. Defaults to ‘green’.

  • bold (bool, optional) – Whether the message is bold. Defaults to False.

Return type:

None

property log_dir: str#

Return the log directory.

register_key(key, window_length=None, min_and_max=False, delta=False)[source]#

Register a key to the logger.

The logger can record the following data:

----------------------------------------------------
|       Name            |            Value         |
----------------------------------------------------
|    Train/Epoch        |             25           |
|  Metrics/EpCost/Min   |            22.38         |
|  Metrics/EpCost/Max   |            25.48         |
|  Metrics/EpCost/Mean  |            23.93         |
----------------------------------------------------
Parameters:

  • key (str) – The name of the key.

  • window_length (int or None, optional) – The length of the window. Defaults to None.

  • min_and_max (bool, optional) – Whether to record the min and max value. Defaults to False.

  • delta (bool, optional) – Whether to record the delta value. Defaults to False.

Return type:

None

save_config(config)[source]#

Save the configuration to the log directory.

Parameters:

config (Config) – The configuration to be saved.

Return type:

None

setup_torch_saver(what_to_save)[source]#

Setup the torch saver.

Parameters:

what_to_save (dict[str, Any]) – The dict of the things to be saved.

Return type:

None

store(data=None, /, **kwargs)[source]#

Store the data to the logger.

Note

The data stored in data will be updated by kwargs.

Parameters:

data (dict[str, float | np.ndarray | torch.Tensor] or None, optional) – The data to be stored. Defaults to None.

Keyword Arguments:

kwargs (int, float, np.ndarray, or torch.Tensor) – The data to be stored.

Return type:

None

torch_save()[source]#

Save the torch model.

Return type:

None