def _plugin_assets(logdir, runs, plugin_name):
result = {}
for run in runs:
run_path = os.path.join(logdir, run)
assets = plugin_asset_util.ListAssets(run_path, plugin_name)
result[run] = assets
return result
def PluginAssets(self, plugin_name):
"""Return a list of all plugin assets for the given plugin.
Args:
plugin_name: The string name of a plugin to retrieve assets for.
Returns:
A list of string plugin asset names, or empty list if none are available.
If the plugin was not registered, an empty list is returned.
"""
return plugin_asset_util.ListAssets(self.path, plugin_name)
def _append_plugin_asset_directories(self, run_path_pairs):
extra = []
plugin_assets_name = metadata.PLUGIN_ASSETS_NAME
for (run, logdir) in run_path_pairs:
assets = plugin_asset_util.ListAssets(logdir, plugin_assets_name)
if metadata.PROJECTOR_FILENAME not in assets:
continue
assets_dir = os.path.join(
self._run_paths[run], metadata.PLUGINS_DIR, plugin_assets_name
)
assets_path_pair = (run, os.path.abspath(assets_dir))
extra.append(assets_path_pair)
run_path_pairs.extend(extra)
def generate_run_to_tools(self):
"""Generator for pairs of "run name" and a list of tools for that run.
The "run name" here is a "frontend run name" - see _run_dir() for the
definition of a "frontend run name" and how it maps to a directory of
profile data for a specific profile "run". The profile plugin concept of
"run" is different from the normal TensorBoard run; each run in this case
represents a single instance of profile data collection, more similar to a
"step" of data in typical TensorBoard semantics. These runs reside in
subdirectories of the plugins/profile directory within any regular
TensorBoard run directory (defined as a subdirectory of the logdir that
contains at least one tfevents file) or within the logdir root directory
itself (even if it contains no tfevents file and would thus not be
considered a normal TensorBoard run, for backwards compatibility).
Within those "profile run directories", there are files in the directory
that correspond to different profiling tools. The file that contains profile
for a specific tool "x" will have a suffix name TOOLS["x"].
Example:
logs/
plugins/
profile/
run1/
hostA.trace
train/
events.out.tfevents.foo
plugins/
profile/
run1/
hostA.trace
hostB.trace
run2/
hostA.trace
validation/
events.out.tfevents.foo
plugins/
profile/
run1/
hostA.trace
Yields:
A sequence of tuples mapping "frontend run names" to lists of tool names
available for those runs. For the above example, this would be:
("run1", ["trace_viewer"])
("train/run1", ["trace_viewer"])
("train/run2", ["trace_viewer"])
("validation/run1", ["trace_viewer"])
"""
self.start_grpc_stub_if_necessary()
plugin_assets = self.multiplexer.PluginAssets(PLUGIN_NAME)
tb_run_names_to_dirs = self.multiplexer.RunPaths()
# Ensure that we also check the root logdir, even if it isn't a recognized
# TensorBoard run (i.e. has no tfevents file directly under it), to remain
# backwards compatible with previously profile plugin behavior. Note that we
# check if logdir is a directory to handle case where it's actually a
# multipart directory spec, which this plugin does not support.
if '.' not in plugin_assets and tf.io.gfile.isdir(self.logdir):
tb_run_names_to_dirs['.'] = self.logdir
plugin_assets['.'] = plugin_asset_util.ListAssets(
self.logdir, PLUGIN_NAME)
for tb_run_name, profile_runs in six.iteritems(plugin_assets):
tb_run_dir = tb_run_names_to_dirs[tb_run_name]
tb_plugin_dir = plugin_asset_util.PluginDirectory(
tb_run_dir, PLUGIN_NAME)
for profile_run in profile_runs:
# Remove trailing separator; some filesystem implementations emit this.
profile_run = profile_run.rstrip(os.sep)
if tb_run_name == '.':
frontend_run = profile_run
else:
frontend_run = os.path.join(tb_run_name, profile_run)
profile_run_dir = os.path.join(tb_plugin_dir, profile_run)
if tf.io.gfile.isdir(profile_run_dir):
yield frontend_run, self._get_active_tools(profile_run_dir)
