wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-08-06 13:31:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

[tune] TensorBoard HParams for TF2.0 (#5678)

Browse Source
This commit is contained in:
Vince Jankovics
2019-09-21 11:06:34 -07:00
committed by Richard Liaw
parent 79b9c70ad6
commit 7e214fd95e
12 changed files with 192 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
doc/source/images/tune-hparams-coord.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 410 KiB

doc/source/images/tune-hparams.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 55 KiB

doc/source/tune-usage.rst
+25 -3
View File
@@ -497,8 +497,8 @@ The following fields will automatically show up on the console output, if provid
Example_0: TERMINATED [pid=68248], 179 s, 2 iter, 60000 ts, 94 rew
Visualizing Results
-------------------
TensorBoard
-----------
To visualize learning in tensorboard, install TensorFlow:
@@ -520,7 +520,28 @@ If you are running Ray on a remote multi-user cluster where you do not have sudo
.. image:: ray-tune-tensorboard.png
To use rllab's VisKit (you may have to install some dependencies), run:
If using TF2, Tune also automatically generates TensorBoard HParams output, as shown below:
.. code-block:: python
tune.run(
...,
config={
"lr": tune.grid_search([1e-5, 1e-4]),
"momentum": tune.grid_search([0, 0.9])
}
)
.. image:: images/tune-hparams.png
The nonrelevant metrics (like timing stats) can be disabled on the left to show only the relevant ones (like accuracy, loss, etc.).
Viskit
------
To use VisKit (you may have to install some dependencies), run:
.. code-block:: bash
@@ -547,6 +568,7 @@ You can pass in your own logging mechanisms to output logs in custom formats as
These loggers will be called along with the default Tune loggers. All loggers must inherit the `Logger interface <tune-package-ref.html#ray.tune.logger.Logger>`__. Tune enables default loggers for Tensorboard, CSV, and JSON formats. You can also check out `logger.py <https://github.com/ray-project/ray/blob/master/python/ray/tune/logger.py>`__ for implementation details. An example can be found in `logging_example.py <https://github.com/ray-project/ray/blob/master/python/ray/tune/examples/logging_example.py>`__.
MLFlow
~~~~~~
doc/source/tune.rst
+5 -1
View File
@@ -1,7 +1,7 @@
Tune: A Scalable Hyperparameter Tuning Library
==============================================
.. important:: Take the 3 minute `2019 Ray Tune User Survey <https://forms.gle/7u5eH1avbTfpZ3dE6>`_!
.. important:: Take the `2019 Ray Tune User Survey <https://forms.gle/7u5eH1avbTfpZ3dE6>`_ and show us your Tune project!
.. image:: images/tune.png
:scale: 30%
@@ -48,6 +48,10 @@ If TensorBoard is installed, automatically visualize all trial results:
.. image:: images/tune-start-tb.png
If using TF2 and TensorBoard, Tune will also automatically generate TensorBoard HParams output:
.. image:: images/tune-hparams-coord.png
Distributed Quick Start
-----------------------
python/ray/tune/logger.py
+25 -7
View File
@@ -38,9 +38,10 @@ class Logger(object):
logdir: Directory for all logger creators to log to.
"""
def __init__(self, config, logdir):
def __init__(self, config, logdir, trial=None):
self.config = config
self.logdir = logdir
self.trial = trial
self._init()
def _init(self):
@@ -135,7 +136,7 @@ class JsonLogger(Logger):
cloudpickle.dump(self.config, f)
def tf2_compat_logger(config, logdir):
def tf2_compat_logger(config, logdir, trial=None):
"""Chooses TensorBoard logger depending on imported TF version."""
global tf
if "RLLIB_TEST_NO_TF_IMPORT" in os.environ:
@@ -148,9 +149,9 @@ def tf2_compat_logger(config, logdir):
distutils.version.LooseVersion("2.0.0"))
if use_tf2_api:
tf = tf.compat.v2 # setting this for TF2.0
return TF2Logger(config, logdir)
return TF2Logger(config, logdir, trial)
else:
return TFLogger(config, logdir)
return TFLogger(config, logdir, trial)
class TF2Logger(Logger):
@@ -166,10 +167,12 @@ class TF2Logger(Logger):
def _init(self):
self._file_writer = None
self._hp_logged = False
def on_result(self, result):
if self._file_writer is None:
from tensorflow.python.eager import context
from tensorboard.plugins.hparams import api as hp
self._context = context
self._file_writer = tf.summary.create_file_writer(self.logdir)
with tf.device("/CPU:0"), self._context.eager_mode():
@@ -178,6 +181,16 @@ class TF2Logger(Logger):
TIMESTEPS_TOTAL) or result[TRAINING_ITERATION]
tmp = result.copy()
if not self._hp_logged:
if self.trial and self.trial.evaluated_params:
try:
hp.hparams(
self.trial.evaluated_params,
trial_id=self.trial.trial_id)
except Exception as exc:
logger.error("HParams failed with %s", exc)
self._hp_logged = True
for k in [
"config", "pid", "timestamp", TIME_TOTAL_S,
TRAINING_ITERATION
@@ -305,7 +318,12 @@ class UnifiedLogger(Logger):
See ray/python/ray/tune/log_sync.py
"""
def __init__(self, config, logdir, loggers=None, sync_function=None):
def __init__(self,
config,
logdir,
trial=None,
loggers=None,
sync_function=None):
if loggers is None:
self._logger_cls_list = DEFAULT_LOGGERS
else:
@@ -313,13 +331,13 @@ class UnifiedLogger(Logger):
self._sync_function = sync_function
self._log_syncer = None
super(UnifiedLogger, self).__init__(config, logdir)
super(UnifiedLogger, self).__init__(config, logdir, trial)
def _init(self):
self._loggers = []
for cls in self._logger_cls_list:
try:
self._loggers.append(cls(self.config, self.logdir))
self._loggers.append(cls(self.config, self.logdir, self.trial))
except Exception as exc:
logger.warning("Could not instantiate {}: {}.".format(
cls.__name__, str(exc)))
python/ray/tune/suggest/basic_variant.py
+4 -3
View File
@@ -8,7 +8,8 @@ import random
from ray.tune.error import TuneError
from ray.tune.experiment import convert_to_experiment_list
from ray.tune.config_parser import make_parser, create_trial_from_spec
from ray.tune.suggest.variant_generator import generate_variants
from ray.tune.suggest.variant_generator import (generate_variants, format_vars,
flatten_resolved_vars)
from ray.tune.suggest.search import SearchAlgorithm
@@ -78,13 +79,13 @@ class BasicVariantGenerator(SearchAlgorithm):
for resolved_vars, spec in generate_variants(unresolved_spec):
experiment_tag = str(self._counter)
if resolved_vars:
experiment_tag += "_{}".format(resolved_vars)
experiment_tag += "_{}".format(format_vars(resolved_vars))
self._counter += 1
yield create_trial_from_spec(
spec,
output_path,
self._parser,
evaluated_params=resolved_vars,
evaluated_params=flatten_resolved_vars(resolved_vars),
experiment_tag=experiment_tag)
def is_finished(self):
python/ray/tune/suggest/suggestion.py
+4 -3
View File
@@ -7,7 +7,7 @@ import copy
from ray.tune.error import TuneError
from ray.tune.trial import Trial
from ray.tune.util import merge_dicts
from ray.tune.util import merge_dicts, flatten_dict
from ray.tune.experiment import convert_to_experiment_list
from ray.tune.config_parser import make_parser, create_trial_from_spec
from ray.tune.suggest.search import SearchAlgorithm
@@ -89,7 +89,8 @@ class SuggestionAlgorithm(SearchAlgorithm):
else:
break
spec = copy.deepcopy(experiment_spec)
spec["config"] = merge_dicts(spec["config"], suggested_config)
spec["config"] = merge_dicts(spec["config"],
copy.deepcopy(suggested_config))
flattened_config = resolve_nested_dict(spec["config"])
self._counter += 1
tag = "{0}_{1}".format(
@@ -98,7 +99,7 @@ class SuggestionAlgorithm(SearchAlgorithm):
spec,
output_path,
self._parser,
evaluated_params=list(suggested_config),
evaluated_params=flatten_dict(suggested_config),
experiment_tag=tag,
trial_id=trial_id)
python/ray/tune/suggest/variant_generator.py
+18 -1
View File
@@ -39,10 +39,15 @@ def generate_variants(unresolved_spec):
"activation": {"grid_search": ["relu", "tanh"]}
"cpu": {"eval": "spec.config.num_workers"}
Use `format_vars` to format the returned dict of hyperparameters.
Yields:
(Dict of resolved variables, Spec object)
"""
for resolved_vars, spec in _generate_variants(unresolved_spec):
assert not _unresolved_values(spec)
yield format_vars(resolved_vars), spec
yield resolved_vars, spec
def grid_search(values):
@@ -79,6 +84,7 @@ def resolve_nested_dict(nested_dict):
def format_vars(resolved_vars):
"""Formats the resolved variable dict into a single string."""
out = []
for path, value in sorted(resolved_vars.items()):
if path[0] in ["run", "env", "resources_per_trial"]:
@@ -96,6 +102,17 @@ def format_vars(resolved_vars):
return ",".join(out)
def flatten_resolved_vars(resolved_vars):
"""Formats the resolved variable dict into a mapping of (str -> value)."""
flattened_resolved_vars_dict = {}
for pieces, value in resolved_vars.items():
if pieces[0] == "config":
pieces = pieces[1:]
pieces = [str(piece) for piece in pieces]
flattened_resolved_vars_dict["/".join(pieces)] = value
return flattened_resolved_vars_dict
def _clean_value(value):
if isinstance(value, float):
return "{:.5}".format(value)
python/ray/tune/tests/test_trial_runner.py
+106 -7
View File
@@ -1202,6 +1202,7 @@ class VariantGeneratorTest(unittest.TestCase):
self.assertEqual(trials[0].trainable_name, "PPO")
self.assertEqual(trials[0].experiment_tag, "0")
self.assertEqual(trials[0].max_failures, 5)
self.assertEqual(trials[0].evaluated_params, {})
self.assertEqual(trials[0].local_dir,
os.path.join(DEFAULT_RESULTS_DIR, "tune-pong"))
self.assertEqual(trials[1].experiment_tag, "1")
@@ -1218,6 +1219,7 @@ class VariantGeneratorTest(unittest.TestCase):
trials = list(trials)
self.assertEqual(len(trials), 1)
self.assertEqual(trials[0].config, {"foo": 4})
self.assertEqual(trials[0].evaluated_params, {"foo": 4})
self.assertEqual(trials[0].experiment_tag, "0_foo=4")
def testGridSearch(self):
@@ -1230,18 +1232,72 @@ class VariantGeneratorTest(unittest.TestCase):
"foo": {
"grid_search": [1, 2, 3]
},
"baz": "asd",
},
}, "grid_search")
trials = list(trials)
self.assertEqual(len(trials), 6)
self.assertEqual(trials[0].config, {"bar": True, "foo": 1})
self.assertEqual(trials[0].config, {
"bar": True,
"foo": 1,
"baz": "asd",
})
self.assertEqual(trials[0].evaluated_params, {
"bar": True,
"foo": 1,
})
self.assertEqual(trials[0].experiment_tag, "0_bar=True,foo=1")
self.assertEqual(trials[1].config, {"bar": False, "foo": 1})
self.assertEqual(trials[1].config, {
"bar": False,
"foo": 1,
"baz": "asd",
})
self.assertEqual(trials[1].evaluated_params, {
"bar": False,
"foo": 1,
})
self.assertEqual(trials[1].experiment_tag, "1_bar=False,foo=1")
self.assertEqual(trials[2].config, {"bar": True, "foo": 2})
self.assertEqual(trials[3].config, {"bar": False, "foo": 2})
self.assertEqual(trials[4].config, {"bar": True, "foo": 3})
self.assertEqual(trials[5].config, {"bar": False, "foo": 3})
self.assertEqual(trials[2].config, {
"bar": True,
"foo": 2,
"baz": "asd",
})
self.assertEqual(trials[2].evaluated_params, {
"bar": True,
"foo": 2,
})
self.assertEqual(trials[3].config, {
"bar": False,
"foo": 2,
"baz": "asd",
})
self.assertEqual(trials[3].evaluated_params, {
"bar": False,
"foo": 2,
})
self.assertEqual(trials[4].config, {
"bar": True,
"foo": 3,
"baz": "asd",
})
self.assertEqual(trials[4].evaluated_params, {
"bar": True,
"foo": 3,
})
self.assertEqual(trials[5].config, {
"bar": False,
"foo": 3,
"baz": "asd",
})
self.assertEqual(trials[5].evaluated_params, {
"bar": False,
"foo": 3,
})
def testGridSearchAndEval(self):
trials = self.generate_trials({
@@ -1250,11 +1306,22 @@ class VariantGeneratorTest(unittest.TestCase):
"qux": tune.sample_from(lambda spec: 2 + 2),
"bar": grid_search([True, False]),
"foo": grid_search([1, 2, 3]),
"baz": "asd",
},
}, "grid_eval")
trials = list(trials)
self.assertEqual(len(trials), 6)
self.assertEqual(trials[0].config, {"bar": True, "foo": 1, "qux": 4})
self.assertEqual(trials[0].config, {
"bar": True,
"foo": 1,
"qux": 4,
"baz": "asd",
})
self.assertEqual(trials[0].evaluated_params, {
"bar": True,
"foo": 1,
"qux": 4,
})
self.assertEqual(trials[0].experiment_tag, "0_bar=True,foo=1,qux=4")
def testConditionResolution(self):
@@ -1269,6 +1336,8 @@ class VariantGeneratorTest(unittest.TestCase):
trials = list(trials)
self.assertEqual(len(trials), 1)
self.assertEqual(trials[0].config, {"x": 1, "y": 2, "z": 3})
self.assertEqual(trials[0].evaluated_params, {"y": 2, "z": 3})
self.assertEqual(trials[0].experiment_tag, "0_y=2,z=3")
def testDependentLambda(self):
trials = self.generate_trials({
@@ -1299,6 +1368,36 @@ class VariantGeneratorTest(unittest.TestCase):
self.assertEqual(trials[0].config, {"x": 100, "y": 1})
self.assertEqual(trials[1].config, {"x": 200, "y": 1})
def testNestedValues(self):
trials = self.generate_trials({
"run": "PPO",
"config": {
"x": {
"y": {
"z": tune.sample_from(lambda spec: 1)
}
},
"y": tune.sample_from(lambda spec: 12),
"z": tune.sample_from(lambda spec: spec.config.x.y.z * 100),
},
}, "nested_values")
trials = list(trials)
self.assertEqual(len(trials), 1)
self.assertEqual(trials[0].config, {
"x": {
"y": {
"z": 1
}
},
"y": 12,
"z": 100
})
self.assertEqual(trials[0].evaluated_params, {
"x/y/z": 1,
"y": 12,
"z": 100
})
def testLogUniform(self):
sampler = tune.loguniform(1e-10, 1e-1).func
results = [sampler(None) for i in range(1000)]
python/ray/tune/trainable.py
+2 -2
View File
@@ -81,8 +81,8 @@ class Trainable(object):
os.makedirs(DEFAULT_RESULTS_DIR)
self._logdir = tempfile.mkdtemp(
prefix=logdir_prefix, dir=DEFAULT_RESULTS_DIR)
self._result_logger = UnifiedLogger(self.config, self._logdir,
None)
self._result_logger = UnifiedLogger(
self.config, self._logdir, loggers=None)
self._iteration = 0
self._time_total = 0.0
python/ray/tune/trial.py
+2 -1
View File
@@ -136,7 +136,7 @@ class Trial(object):
self.local_dir = local_dir # This remains unexpanded for syncing.
#: Parameters that Tune varies across searches.
self.evaluated_params = evaluated_params or []
self.evaluated_params = evaluated_params or {}
self.experiment_tag = experiment_tag
trainable_cls = self._get_trainable_cls()
if trainable_cls and hasattr(trainable_cls,
@@ -237,6 +237,7 @@ class Trial(object):
self.result_logger = UnifiedLogger(
self.config,
self.logdir,
trial=self,
loggers=self.loggers,
sync_function=self.sync_to_driver_fn)
rllib/agents/trainer.py
+1 -1
View File
@@ -365,7 +365,7 @@ class Trainer(Trainable):
os.makedirs(DEFAULT_RESULTS_DIR)
logdir = tempfile.mkdtemp(
prefix=logdir_prefix, dir=DEFAULT_RESULTS_DIR)
return UnifiedLogger(config, logdir, None)
return UnifiedLogger(config, logdir, loggers=None)
logger_creator = default_logger_creator