def testInvalidVerbosityRaisesError(self):
with self.assertRaises(ValueError):
upload_tracker.UploadTracker(verbosity="1")
with self.assertRaises(ValueError):
upload_tracker.UploadTracker(verbosity=-1)
with self.assertRaises(ValueError):
upload_tracker.UploadTracker(verbosity=0.5)
with self.assertRaises(ValueError):
upload_tracker.UploadTracker(verbosity=100)
with self.assertRaises(ValueError):
upload_tracker.UploadTracker(verbosity=None)
def testBlobTrackerWithVerbosity0(self):
tracker = upload_tracker.UploadTracker(verbosity=0)
with tracker.blob_tracker(blob_bytes=2048):
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
def testBlobTrackerNotUploaded(self):
tracker = upload_tracker.UploadTracker(verbosity=1)
with tracker.send_tracker():
self.assertEqual(self.mock_write.call_count, 2)
self.assertEqual(self.mock_flush.call_count, 2)
self.assertIn(
"Started scanning",
self.mock_write.call_args_list[0][0][0],
)
with tracker.blob_tracker(blob_bytes=2048 * 1024 *
1024) as blob_tracker:
self.assertEqual(self.mock_write.call_count, 3)
self.assertEqual(self.mock_flush.call_count, 3)
self.assertIn(
"Uploading binary object (2048.0 MB)",
self.mock_write.call_args[0][0],
)
blob_tracker.mark_uploaded(is_uploaded=False)
self.assertEqual(self.mock_write.call_count, 6)
self.assertEqual(self.mock_flush.call_count, 5)
self.assertIn(
"Total uploaded: 0 scalars, 0 tensors, 0 binary objects\n",
self.mock_write.call_args_list[3][0][0],
)
self.assertIn(
"Total skipped: 1 binary objects (2048.0 MB)\n",
self.mock_write.call_args_list[4][0][0],
)
self.assertEqual(tracker.has_data(), True)
def testScalarsTrackerWithVerbosity0(self):
tracker = upload_tracker.UploadTracker(verbosity=0)
with tracker.scalars_tracker(123):
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
def testBlobTrackerUploaded(self):
tracker = upload_tracker.UploadTracker(verbosity=1)
with tracker.blob_tracker(blob_bytes=2048) as blob_tracker:
self.assertEqual(self.mock_write.call_count, 1)
self.assertEqual(self.mock_flush.call_count, 1)
self.assertIn(
"Uploading binary object (2.0 kB)",
self.mock_write.call_args[0][0],
)
def testTensorsTrackerWithVerbosity0(self):
tracker = upload_tracker.UploadTracker(verbosity=0)
with tracker.tensors_tracker(
num_tensors=200,
num_tensors_skipped=50,
tensor_bytes=6000,
tensor_bytes_skipped=4000,
):
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
self.assertEqual(self.mock_write.call_count, 0)
self.assertEqual(self.mock_flush.call_count, 0)
def testScalarsTracker(self):
tracker = upload_tracker.UploadTracker(verbosity=1)
with tracker.scalars_tracker(123):
self.assertEqual(self.mock_write.call_count, 1)
self.assertEqual(self.mock_flush.call_count, 1)
self.assertIn(
"Uploading 123 scalars...",
self.mock_write.call_args[0][0],
)
self.assertEqual(self.mock_write.call_count, 1)
self.assertEqual(self.mock_flush.call_count, 1)
self.assertEqual(tracker.has_data(), True)
def testTensorsTrackerWithoutSkippedTensors(self):
tracker = upload_tracker.UploadTracker(verbosity=1)
with tracker.tensors_tracker(
num_tensors=200,
num_tensors_skipped=0,
tensor_bytes=6000,
tensor_bytes_skipped=0,
):
self.assertEqual(self.mock_write.call_count, 1)
self.assertEqual(self.mock_flush.call_count, 1)
self.assertIn(
"Uploading 200 tensors (5.9 kB)",
self.mock_write.call_args[0][0],
)
self.assertEqual(tracker.has_data(), True)
def testSendTracker(self):
tracker = upload_tracker.UploadTracker(verbosity=1)
with tracker.send_tracker():
self.assertEqual(self.mock_write.call_count, 2)
self.assertEqual(self.mock_flush.call_count, 2)
self.assertIn(
"Data upload starting...",
self.mock_write.call_args[0][0],
)
self.assertEqual(self.mock_write.call_count, 3)
self.assertEqual(self.mock_flush.call_count, 3)
self.assertIn(
"Listening for new data in logdir...",
self.mock_write.call_args[0][0],
)
self.assertEqual(tracker.has_data(), False)
def create_experiment(self):
"""Creates an Experiment for this upload session and returns the ID."""
logger.info("Creating experiment")
request = write_service_pb2.CreateExperimentRequest(
name=self._name, description=self._description)
response = grpc_util.call_with_retries(self._api.CreateExperiment,
request)
self._tracker = upload_tracker.UploadTracker(verbosity=self._verbosity)
self._request_sender = _BatchedRequestSender(
response.experiment_id,
self._api,
allowed_plugins=self._allowed_plugins,
upload_limits=self._upload_limits,
rpc_rate_limiter=self._rpc_rate_limiter,
tensor_rpc_rate_limiter=self._tensor_rpc_rate_limiter,
blob_rpc_rate_limiter=self._blob_rpc_rate_limiter,
tracker=self._tracker,
)
return response.experiment_id
def create_profile_request_sender() -> profile_uploader.ProfileRequestSender:
"""Creates the `ProfileRequestSender` for the profile plugin.
A profile request sender is created for the plugin so that after profiling runs
have finished, data can be uploaded to the tensorboard backend.
Returns:
A ProfileRequestSender object.
"""
api_client = _get_api_client()
experiment_name = _get_or_create_experiment(
api_client, training_utils.environment_variables.cloud_ml_job_id)
upload_limits = _make_upload_limits()
blob_rpc_rate_limiter = util.RateLimiter(
upload_limits.min_blob_request_interval / 100)
blob_storage_bucket, blob_storage_folder = _get_blob_items(api_client, )
source_bucket = uploader_utils.get_source_bucket(
training_utils.environment_variables.tensorboard_log_dir)
profile_request_sender = profile_uploader.ProfileRequestSender(
experiment_name,
api_client,
upload_limits=upload_limits,
blob_rpc_rate_limiter=blob_rpc_rate_limiter,
blob_storage_bucket=blob_storage_bucket,
blob_storage_folder=blob_storage_folder,
source_bucket=source_bucket,
tracker=upload_tracker.UploadTracker(verbosity=1),
logdir=training_utils.environment_variables.tensorboard_log_dir,
)
return profile_request_sender
def __init__(
self,
writer_client,
logdir,
allowed_plugins,
upload_limits,
logdir_poll_rate_limiter=None,
rpc_rate_limiter=None,
tensor_rpc_rate_limiter=None,
blob_rpc_rate_limiter=None,
name=None,
description=None,
verbosity=None,
one_shot=None,
):
"""Constructs a TensorBoardUploader.
Args:
writer_client: a TensorBoardWriterService stub instance
logdir: path of the log directory to upload
allowed_plugins: collection of string plugin names; events will only
be uploaded if their time series's metadata specifies one of these
plugin names
upload_limits: instance of tensorboard.service.UploadLimits proto.
logdir_poll_rate_limiter: a `RateLimiter` to use to limit logdir
polling frequency, to avoid thrashing disks, especially on networked
file systems
rpc_rate_limiter: a `RateLimiter` to use to limit write RPC frequency.
Note this limit applies at the level of single RPCs in the Scalar
and Tensor case, but at the level of an entire blob upload in the
Blob case-- which may require a few preparatory RPCs and a stream
of chunks. Note the chunk stream is internally rate-limited by
backpressure from the server, so it is not a concern that we do not
explicitly rate-limit within the stream here.
name: String name to assign to the experiment.
description: String description to assign to the experiment.
verbosity: Level of verbosity, an integer. Supported value:
0 - No upload statistics is printed.
1 - Print upload statistics while uploading data (default).
one_shot: Once uploading starts, upload only the existing data in
the logdir and then return immediately, instead of the default
behavior of continuing to listen for new data in the logdir and
upload them when it appears.
"""
self._api = writer_client
self._logdir = logdir
self._allowed_plugins = frozenset(allowed_plugins)
self._upload_limits = upload_limits
self._name = name
self._description = description
self._verbosity = 1 if verbosity is None else verbosity
self._one_shot = False if one_shot is None else one_shot
self._request_sender = None
self._experiment_id = None
if logdir_poll_rate_limiter is None:
self._logdir_poll_rate_limiter = util.RateLimiter(
_MIN_LOGDIR_POLL_INTERVAL_SECS)
else:
self._logdir_poll_rate_limiter = logdir_poll_rate_limiter
if rpc_rate_limiter is None:
self._rpc_rate_limiter = util.RateLimiter(
self._upload_limits.min_scalar_request_interval / 1000)
else:
self._rpc_rate_limiter = rpc_rate_limiter
if tensor_rpc_rate_limiter is None:
self._tensor_rpc_rate_limiter = util.RateLimiter(
self._upload_limits.min_tensor_request_interval / 1000)
else:
self._tensor_rpc_rate_limiter = tensor_rpc_rate_limiter
if blob_rpc_rate_limiter is None:
self._blob_rpc_rate_limiter = util.RateLimiter(
self._upload_limits.min_blob_request_interval / 1000)
else:
self._blob_rpc_rate_limiter = blob_rpc_rate_limiter
active_filter = (
lambda secs: secs + _EVENT_FILE_INACTIVE_SECS >= time.time())
directory_loader_factory = functools.partial(
directory_loader.DirectoryLoader,
loader_factory=event_file_loader.TimestampedEventFileLoader,
path_filter=io_wrapper.IsTensorFlowEventsFile,
active_filter=active_filter,
)
self._logdir_loader = logdir_loader.LogdirLoader(
self._logdir, directory_loader_factory)
self._tracker = upload_tracker.UploadTracker(verbosity=self._verbosity,
one_shot=self._one_shot)
