hyrax.tensorboardx_logger

Attributes

tensorboardx_logger
tensorboard_start_ns

Classes

HyraxSummaryWriter

This is a wrapper class around TensorboardX SummaryWriter that allows definition

Functions

get_tensorboard_logger()	Get a Tensorboard logging object.
init_tensorboard_logger(**kwargs)	Initialize a Tensorboard SummaryWriter for use by the whole process.
close_tensorboard_logger()	Close the existing global tensorboard logger.

Module Contents

tensorboardx_logger = None

tensorboard_start_ns = 0

get_tensorboard_logger()

Get a Tensorboard logging object.

This is a way for code that just needs to log data to get a handle to the global tensorboard logger. Client code does not need to handle initialization order, it can simply log.

If there is no tensorboard logger set up, messages will be dropped silently.

init_tensorboard_logger(**kwargs)

Initialize a Tensorboard SummaryWriter for use by the whole process.

Args are those to SummaryWriter’s constructor.

If there already is a Tensorboard Logger initialized for this process, it will be closed by this function call before the new one is initialized.

This should be called by code that controls overall hyrax execution e.g. a Verb’s run() method.

close_tensorboard_logger()

Close the existing global tensorboard logger.

If there is no global tensorboard logger, this does nothing.

class HyraxSummaryWriter

This is a wrapper class around TensorboardX SummaryWriter that allows definition of convenience methods for commonly-used logging.

For client code that just wants to log Typical usage is:

from hyrax.tensorboardx_logger import get_tensorboard_logger tensorboardx_logger = get_tensorboard_logger()

…in code…

tensorboardx_logger.log_scalar(…) tensorboardx_logger.log_duration(…)

For code controlling overall process execution (e.g. a hyrax verb run method) usage looks like:

from hyrax.tensorboardx_logger import init_tensorboard_logger, close_tensorboard_logger

def run():

init_tensorboard_logger(log_dir=”some/path/”)

… do things …

close_tensorboard_logger() return

__dir__ and __getattr__ pass through function calls to the underlying tensorboardX SummaryWriter if it exists. Otherwise empty/noop objects are returned. We don’t use inheritance here because we want consumers to not have to think about initialization order concerns, yet have a handle to a pile of functions that all log to the one true tensorboard instance (if it exists)

All functions defined here need to be noops when global tensorboardx_logger is None.

We have the capacity to place information on instances of this class (e.g. a name prefix) but its not implemented. One major issue is providing continuity of interface with tensorboard’s functions that don’t recognize a name prefix. For now, the fully qualified tensorboard name of the data is the common interface, since that follows what tensorboard functions expect.

log_duration_ts(name: str, start_time: int)

Log a duration to tensorboardX as a time series if configured.

Caller provides the start of the duration in time.monotonic_ns End of the duration is assumed to be the moment the function is called.

Parameters:

• name (str) – The name of the scalar to log

• start_time (int) – Start time in nanoseconds from time.monotonic_ns()

log_scalar_ts(name: str, scalar, since_tensorboard_start_ns=None)

Log a scalar to tensorboardX as a time series if configured.

Parameters:

• name (str) – The name of the scalar to log

• scalar (Any) – The value to log. Really ought to be a number.

• since_tensorboard_start_ns (int, Optional) – Log time in nanoseconds from the beginning of tensorboard logging. If not provided, this will be calculated at the moment this function is called

__dir__()

__getattr__(name)