def test_atomic_read(self, name):
dataframe = pd.DataFrame(dict(a=[1, 2], b=[4.0, 5.0]))
csv_file = os.path.join(absltest.get_default_test_tmpdir(), name)
utils_impl.atomic_write_to_csv(dataframe, csv_file)
dataframe2 = utils_impl.atomic_read_from_csv(csv_file)
pd.testing.assert_frame_equal(dataframe, dataframe2)
def test_rows_are_cleared_is_reflected_in_saved_file(self):
temp_dir = self.get_temp_dir()
metrics_mngr = metrics_manager.ScalarMetricsManager(temp_dir,
prefix='foo')
metrics_mngr.update_metrics(0, _create_dummy_metrics())
metrics_mngr.update_metrics(5, _create_dummy_metrics())
metrics_mngr.update_metrics(10, _create_dummy_metrics())
file_contents_before = utils_impl.atomic_read_from_csv(
os.path.join(temp_dir, 'foo.metrics.csv.bz2'))
self.assertEqual(3, len(file_contents_before.index))
metrics_mngr.clear_rounds_after(last_valid_round_num=7)
file_contents_after = utils_impl.atomic_read_from_csv(
os.path.join(temp_dir, 'foo.metrics.csv.bz2'))
self.assertEqual(2, len(file_contents_after.index))
def __init__(self,
root_metrics_dir: str = '/tmp',
prefix: str = 'experiment',
use_bz2: bool = True):
"""Returns an initialized `ScalarMetricsManager`.
This class will maintain metrics in a CSV file in the filesystem. The path
of the file is {`root_metrics_dir`}/{`prefix`}.metrics.csv (if use_bz2 is
set to False) or {`root_metrics_dir`}/{`prefix`}.metrics.csv.bz2 (if
use_bz2 is set to True). To use this class upon restart of an experiment at
an earlier round number, you can initialize and then call the
clear_rounds_after() method to remove all rows for round numbers later than
the restart round number. This ensures that no duplicate rows of data exist
in the CSV.
Args:
root_metrics_dir: A path on the filesystem to store CSVs.
prefix: A string to use as the prefix of filename. Usually the name of a
specific run in a larger grid of experiments sharing a common
`root_metrics_dir`.
use_bz2: A boolean indicating whether to zip the result metrics csv using
bz2.
Raises:
ValueError: If `root_metrics_dir` is empty string.
ValueError: If `prefix` is empty string.
ValueError: If the specified metrics csv file already exists but does not
contain a `round_num` column.
"""
super().__init__()
if not root_metrics_dir:
raise ValueError(
'Empty string passed for root_metrics_dir argument.')
if not prefix:
raise ValueError('Empty string passed for prefix argument.')
if use_bz2:
# Using .bz2 rather than .zip due to
# https://github.com/pandas-dev/pandas/issues/26023
self._metrics_filename = os.path.join(root_metrics_dir,
f'{prefix}.metrics.csv.bz2')
else:
self._metrics_filename = os.path.join(root_metrics_dir,
f'{prefix}.metrics.csv')
if not tf.io.gfile.exists(self._metrics_filename):
utils_impl.atomic_write_to_csv(pd.DataFrame(),
self._metrics_filename)
self._metrics = utils_impl.atomic_read_from_csv(self._metrics_filename)
if not self._metrics.empty and 'round_num' not in self._metrics.columns:
raise ValueError(
f'The specified csv file ({self._metrics_filename}) already exists '
'but was not created by ScalarMetricsManager (it does not contain a '
'`round_num` column.')
self._latest_round_num = (None if self._metrics.empty else
self._metrics.round_num.max(axis=0))
def on_epoch_end(self, epoch: int, logs: Dict[Any, Any] = None):
results_path = os.path.join(self._path, 'metric_results.csv')
if tf.io.gfile.exists(results_path):
# Read the results until now.
results_df = utils_impl.atomic_read_from_csv(results_path)
# Slice off results after the current epoch, this indicates the job
# restarted.
results_df = results_df[:epoch]
# Add the new epoch.
results_df = results_df.append(logs, ignore_index=True)
else:
results_df = pd.DataFrame(logs, index=[epoch])
utils_impl.atomic_write_to_csv(results_df, results_path)