class composer.loggers.TensorboardLogger(log_dir=None, flush_interval=100, rank_zero_only=True)[source]#

Log to Tensorboard.

If you are accessing your logs from a cloud bucket, like S3, they will be in {your_bucket_name}/tensorboard_logs/{run_name} with names like events.out.tfevents-{run_name}-{rank}.

If you are accessing your logs locally (from wherever you are running composer), the logs will be in the relative path: tensorboard_logs/{run_name} with names starting with events.out.tfevents.*

  • log_dir (str, optional) โ€“ The path to the directory where all the tensorboard logs will be saved. This is also the value that should be specified when starting a tensorboard server. e.g. tensorboard โ€“logdir={log_dir}. If not specified ./tensorboard_logs will be used.

  • flush_interval (int, optional) โ€“ How frequently by batch to flush the log to a file. For example, a flush interval of 10 means the log will be flushed to a file every 10 batches. The logs will also be automatically flushed at the start and end of every evaluation phase (EVENT.EVAL_START and EVENT.EVAL_END ), the end of every epoch (EVENT.EPOCH_END), and the end of training (EVENT.FIT_END). Default: 100.

  • rank_zero_only (bool, optional) โ€“ Whether to log only on the rank-zero process. Recommended to be true since the rank 0 will have access to most global metrics. A setting of False may lead to logging of duplicate values. Default: True.