vak.common.tensorboard.events2df¶
- vak.common.tensorboard.events2df(events_path: str | Path, size_guidance: dict | None = None, drop_wall_time: bool = True) DataFrame [source]¶
Convert
tensorboard
events file to pandas.DataFrameEvents files are created by SummaryWriter from PyTorch or Tensorflow.
- Parameters:
events_path (str, pathlib.Path) – Path to either a log directory or a specific events file saved by a SummaryWriter in a log directory. By default,
vak
saves logs in a directory with the model name inside aresults
directory generated at the start of training.size_guidance (dict) – Argument passed to the
EventAccumlator
class fromtensorboard
that is used to load the events file. Information on how much data the EventAccumulator should store in memory. Dict that maps a tagType string to an integer representing the number of items to keep per tag for items of that tagType. If the size is 0, all events are stored. Default is None, in which casevak.tensorboard.DEFAULT_SIZE_GUIDANCE
is used. For more information see https://github.com/tensorflow/tensorboard/blob/master/tensorboard/backend/event_processing/event_accumulator.pydrop_wall_time (bool) – If True, drop wall times logged in events file. Default is True.
- Returns:
df – With index ‘step’ and all Scalars from the events file
- Return type:
pandas.Dataframe
Examples
>>> events_path = 'tweetynet/results_210322_103904/train_dur_6s/replicate_2/TweetyNet/' >>> events_df = vak.tensorboard.events2df(events_path) >>> events_df loss/train avg_acc/val avg_levenshtein/val avg_character_error_rate/val avg_loss/val step 0 2.479142 NaN NaN NaN NaN 1 2.458833 NaN NaN NaN NaN 2 2.441571 NaN NaN NaN NaN 3 2.402737 NaN NaN NaN NaN 4 2.404369 NaN NaN NaN NaN ... ... ... ... ... ... 996 0.171681 NaN NaN NaN NaN 997 0.100202 NaN NaN NaN NaN 998 0.073055 NaN NaN NaN NaN 999 0.031479 NaN NaN NaN NaN 1000 NaN 0.902475 42.0 0.880533 0.310385
[1001 rows x 5 columns]