Tracker

Here is a tutorial on Google Colab that shows how to use the tracker module

Track

labml.tracker.save()

labml.tracker.save(global_step: int)

labml.tracker.save(values: Dict[str, any])

labml.tracker.save(name: str, value: any)

labml.tracker.save(**kwargs: any)

labml.tracker.save(global_step: int, values: Dict[str, any])

labml.tracker.save(global_step: int, name: str, value: any)

labml.tracker.save(global_step: int, **kwargs: any)

This has multiple overloads

labml.tracker.save()

labml.tracker.save(global_step: int)

labml.tracker.save(values: Dict[str, any])

labml.tracker.save(name: str, value: any)

labml.tracker.save(**kwargs: any)

labml.tracker.save(global_step: int, values: Dict[str, any])

labml.tracker.save(global_step: int, name: str, value: any)

labml.tracker.save(global_step: int, **kwargs: any)

This saves the tracking information in all the writers such as labml.ai monitoring app, TensorBoard and Weights and Biases.

Parameters

• global_step (int) – The current step

• values (Dict[str, any]) – A dictionary of key-value pairs to track

• name (str) – The name of the value to be tracked

• value (any) – The value to be tracked

• kwargs – Key-value pairs to track

labml.tracker.add(values: Dict[str, any])

labml.tracker.add(name: str, value: any)

labml.tracker.add(**kwargs: any)

This has multiple overloads

labml.tracker.add(values: Dict[str, any])

labml.tracker.add(name: str, value: any)

labml.tracker.add(**kwargs: any)

This add tracking information to a temporary queue. These will be saved when labml.tracker.save() is called.

You should use labml.tracker.add() to improve performance since saving tracking information consumes time. Although saving takes negligible amount of time it can add up if called very frequently.

Parameters

• values (Dict[str, any]) – A dictionary of key-value pairs to track

• name (str) – The name of the value to be tracked

• value (any) – The value to be tracked

• kwargs – Key-value pairs to track

Step

labml.tracker.set_global_step(global_step: Optional[int])

Set the current step for tracking

Parameters

global_step (int) – Global step

labml.tracker.add_global_step(increment_global_step: int = 1)

Increment the current step for tracking

Parameters

increment_global_step (int, optional) – By how much to increment the global step. Defaults to 1 if not provided.

labml.tracker.get_global_step() → int

Returns current step

Setup

labml.tracker.set_queue(name: str, queue_size: int = 10, is_print: bool = False)

Set indicator type to be a queue. This will maintain a queue of size queue_size to store the tracked values. A histogram of the queue contents and stats like mean will be logged.

This is useful when we want to track statistics like moving average.

Parameters

• name (str) – Name of the indicator

• queue_size (int, optional) – Size of the queue. Defaults to 10.

• is_print – (bool, optional): Whether the indicator should be printed in console. Defaults to False.

labml.tracker.set_histogram(name: str, is_print: bool = False)

Set indicator type to be a histogram. It will log the tracked values as a histogram.

Parameters

• name (str) – Name of the indicator

• is_print – (bool, optional): Whether the indicator should be printed in console. Defaults to False.

labml.tracker.set_scalar(name: str, is_print: bool = False)

Set indicator type to be a scalar. It will log a scalar of the tracked values. If there are multiple values it will log the mean.

Parameters

• name (str) – Name of the indicator

• is_print – (bool, optional): Whether the indicator should be printed in console. Defaults to False.

labml.tracker.set_indexed_scalar(name: str)

Set indicator type to be an indexed scalar. It will log pairs of values (index, value).

Parameters

name (str) – Name of the indicator

Warning

Artificact setup functions labml.tracker.set_image(), labml.tracker.set_text(), labml.tracker.set_tensor(), and labml.tracker.set_indexed_text() are still experimental.

labml.tracker.set_image(name: str, is_print: bool = False, density: Optional[float] = None)

Set indicator type to be an image.

Parameters

• name (str) – Name of the indicator

• is_print – (bool, optional): Whether to show the image with matplotlib. Defaults to False.

• density – (float, optional): This controls how often to log images.

labml.tracker.set_text(name: str, is_print: bool = False)

Set indicator type to be text (a string).

Parameters

• name (str) – Name of the indicator

• is_print – (bool, optional): Whether to show the image with matplotlib. Defaults to False.

labml.tracker.set_tensor(name: str, is_once: bool = False)

Set indicator type to be a tensor.

Parameters

• name (str) – Name of the indicator

• is_once – (bool, optional): Whether this is tracked once only

labml.tracker.set_indexed_text(name: str, title: Optional[str] = None, is_print: bool = False)

Set indicator type to be an indexed text. It will log (index, text) pairs.

Parameters

• name (str) – Name of the indicator

• title (str) – Title to display

• is_print – (bool, optional): Whether to show the image with matplotlib. Defaults to False.

Namespaces

labml.tracker.namespace(name: str)

Set a namespace for tracking

Helpers

labml.tracker.reset()

Reset indicators, for a new experiment

labml.tracker.new_line()

Prints a new line.

Equivalent to logger.log, but handles distributed training where only the rank=0 process is tracking data.