core.common.logger

Copyright (c) Meta, Inc. and its affiliates.

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

Classes

Logger	Generic class to interface with various logging modules, e.g. wandb,
WandBLogger	Generic class to interface with various logging modules, e.g. wandb,
TensorboardLogger	Generic class to interface with various logging modules, e.g. wandb,
WandBSingletonLogger	Singleton version of wandb logger, this forces a single instance of the logger to be created and used from anywhere in the code (not just from the trainer).

Functions

log_stats(x, prefix)

Module Contents

class core.common.logger.Logger(config)

Bases: abc.ABC

Generic class to interface with various logging modules, e.g. wandb, tensorboard, etc.

config

abstract watch(model, log_freq: int = 1000)

Monitor parameters and gradients.

log(update_dict, step: int, split: str = '')

Log some values.

abstract log_plots(plots) → None

abstract mark_preempting() → None

abstract log_summary(summary_dict: dict[str, Any]) → None

abstract log_artifact(name: str, type: str, file_location: str) → None

class core.common.logger.WandBLogger(config)

Bases: Logger

Generic class to interface with various logging modules, e.g. wandb, tensorboard, etc.

watch(model, log='all', log_freq: int = 1000) → None

Monitor parameters and gradients.

log(update_dict, step: int, split: str = '') → None

Log some values.

log_plots(plots, caption: str = '') → None

log_table(name: str, cols: list, data: list, step: int | None = None, commit=False) → None

log_summary(summary_dict: dict[str, Any])

mark_preempting() → None

log_artifact(name: str, type: str, file_location: str) → None

class core.common.logger.TensorboardLogger(config)

Bases: Logger

Generic class to interface with various logging modules, e.g. wandb, tensorboard, etc.

writer

watch(model, log_freq: int = 1000) → bool

Monitor parameters and gradients.

log(update_dict, step: int, split: str = '')

Log some values.

mark_preempting() → None

log_plots(plots) → None

log_summary(summary_dict: dict[str, Any]) → None

log_artifact(name: str, type: str, file_location: str) → None

class core.common.logger.WandBSingletonLogger

Singleton version of wandb logger, this forces a single instance of the logger to be created and used from anywhere in the code (not just from the trainer). This will replace the original WandBLogger.

We initialize wandb instance somewhere in the trainer/runner globally:

WandBSingletonLogger.init_wandb(…)

Then from anywhere in the code we can fetch the singleton instance and log to wandb, note this allows you to log without knowing explicitly which step you are on see: https://docs.wandb.ai/ref/python/log/#the-wb-step for more details

WandBSingletonLogger.get_instance().log({“some_value”: value}, commit=False)

_instance = None

classmethod initialized() → bool

classmethod init_wandb(config: dict, run_id: str, run_name: str, log_dir: str, project: str, entity: str, group: str | None = None) → None

classmethod get_instance()

watch(model, log='all', log_freq: int = 1000) → None

log(update_dict: dict, step: int | None = None, commit=False, split: str = '') → None

log_table(name: str, cols: list, data: list, step: int | None = None, commit=False) → None

log_summary(summary_dict: dict[str, Any])

mark_preempting() → None

log_artifact(name: str, type: str, file_location: str) → None

core.common.logger.log_stats(x: torch.Tensor, prefix: str)