Logger and LogLevel

class composer.core.logging.logger.LogLevel(value)[source]

LogLevel denotes where in the training loop log messages are generated.

Logging destinations use the LogLevel to determine whether to record a given metric or state change.

FIT: Logged once per training run.

EPOCH: Logged once per epoch.

BATCH: Logged once per batch.

MICROBATCH: Logged once per microbatch (e.g. forward pass).

VERBOSE: Logged for debugging.

class composer.core.logging.logger.Logger(state: State, backends: Sequence[BaseLoggerBackend] = ())[source]

Logger routes metrics to the BaseLoggerBackend.

Parameters

state (State) – The global State object.
backends (Sequence[BaseLoggerBackend]) – A sequence of BaseLoggerBackends to which logging calls will be sent.

backends

A sequence of BaseLoggerBackends to which logging calls will be sent.

Type: Sequence[BaseLoggerBackend]

metric(log_level: Union[str, composer.core.logging.logger.LogLevel], data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]

Log a metric to the backends.

Parameters

log_level (Union[str, LogLevel]) – A LogLevel.
data (Union[TLogData, Callable[[], TLogData]]) – Can be either logging data or a callable that returns data to be logged. Callables will be invoked only when will_log() returns True for at least one BaseLoggerBackend. Useful when it is expensive to generate the data to be logged.

metric_batch(data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]: Helper function for self.metric(LogLevel.BATCH, data)

metric_epoch(data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]: Helper function for self.metric(LogLevel.EPOCH, data)

metric_fit(data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]: Helper function for metric(LogLevel.FIT, data)

metric_microbatch(data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]: Helper function for self.metric(LogLevel.MICROBATCH, data)

metric_verbose(data: Union[Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]], Callable[[], Mapping[str, Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]]]]]) → None[source]: Helper function for self.metric(LogLevel.VERBOSE, data)

composer.core.logging.logger.format_log_data_value(data: Union[str, float, int, torch.Tensor, Sequence[TLogDataValue], Mapping[str, TLogDataValue]], /) → str[source]

Recursively formats a given log data value into a string.

Parameters: data – Data to format.
Returns: str – The data, as a string.