From c9740eb5a9a0abac7ce555a51db162959a33f77c Mon Sep 17 00:00:00 2001
From: hershg <hersh.godse@gmail.com>
Date: Fri, 24 May 2019 22:09:14 -0700
Subject: [PATCH] refactored code to use metric,mode instead of reward_attr for
 schedulers (#4120)

---
 ci/long_running_tests/workloads/pbt.py                  | 3 ++-
 doc/source/tune-schedulers.rst                          | 8 +++++---
 python/ray/tune/examples/async_hyperband_example.py     | 3 ++-
 python/ray/tune/examples/ax_example.py                  | 2 +-
 python/ray/tune/examples/bayesopt_example.py            | 2 +-
 python/ray/tune/examples/genetic_example.py             | 2 +-
 python/ray/tune/examples/hyperband_example.py           | 3 ++-
 python/ray/tune/examples/hyperopt_example.py            | 2 +-
 python/ray/tune/examples/mnist_pytorch.py               | 3 ++-
 python/ray/tune/examples/mnist_pytorch_trainable.py     | 2 +-
 python/ray/tune/examples/nevergrad_example.py           | 2 +-
 python/ray/tune/examples/pbt_example.py                 | 3 ++-
 python/ray/tune/examples/pbt_ppo_example.py             | 3 ++-
 python/ray/tune/examples/pbt_tune_cifar10_with_keras.py | 3 ++-
 python/ray/tune/examples/sigopt_example.py              | 2 +-
 python/ray/tune/examples/skopt_example.py               | 4 ++--
 python/ray/tune/examples/tune_cifar10_gluon.py          | 3 ++-
 python/ray/tune/examples/tune_mnist_async_hyperband.py  | 3 ++-
 python/ray/tune/examples/tune_mnist_keras.py            | 3 ++-
 python/ray/tune/examples/tune_mnist_ray_hyperband.py    | 2 +-
 python/ray/tune/tests/test_experiment_analysis.py       | 3 ++-
 python/ray/tune/tests/test_trial_scheduler.py           | 8 +++++---
 22 files changed, 42 insertions(+), 27 deletions(-)

diff --git a/ci/long_running_tests/workloads/pbt.py b/ci/long_running_tests/workloads/pbt.py
index 86473d86ec4d..5e63596c4a79 100644
--- a/ci/long_running_tests/workloads/pbt.py
+++ b/ci/long_running_tests/workloads/pbt.py
@@ -37,7 +37,8 @@
 
 pbt = PopulationBasedTraining(
     time_attr="training_iteration",
-    reward_attr="episode_reward_mean",
+    metric="episode_reward_mean",
+    mode="max",
     perturbation_interval=10,
     hyperparam_mutations={
         "lr": [0.1, 0.01, 0.001, 0.0001],
diff --git a/doc/source/tune-schedulers.rst b/doc/source/tune-schedulers.rst
index cbb105ff5806..f9734181eb37 100644
--- a/doc/source/tune-schedulers.rst
+++ b/doc/source/tune-schedulers.rst
@@ -7,7 +7,7 @@ By default, Tune schedules trials in serial order with the ``FIFOScheduler`` cla
 
     tune.run( ... , scheduler=AsyncHyperBandScheduler())
 
-Tune includes distributed implementations of early stopping algorithms such as `Median Stopping Rule <https://research.google.com/pubs/pub46180.html>`__, `HyperBand <https://arxiv.org/abs/1603.06560>`__, and an `asynchronous version of HyperBand <https://openreview.net/forum?id=S1Y7OOlRZ>`__. These algorithms are very resource efficient and can outperform Bayesian Optimization methods in `many cases <https://people.eecs.berkeley.edu/~kjamieson/hyperband.html>`__. Currently, all schedulers take in a ``reward_attr``, which is assumed to be maximized.
+Tune includes distributed implementations of early stopping algorithms such as `Median Stopping Rule <https://research.google.com/pubs/pub46180.html>`__, `HyperBand <https://arxiv.org/abs/1603.06560>`__, and an `asynchronous version of HyperBand <https://openreview.net/forum?id=S1Y7OOlRZ>`__. These algorithms are very resource efficient and can outperform Bayesian Optimization methods in `many cases <https://people.eecs.berkeley.edu/~kjamieson/hyperband.html>`__. Currently, all schedulers take in a ``metric``, which is minimized or maximized according to ``mode``.
 
 Current Available Trial Schedulers:
 
@@ -25,7 +25,8 @@ Tune includes a distributed implementation of `Population Based Training (PBT) <
 
     pbt_scheduler = PopulationBasedTraining(
             time_attr='time_total_s',
-            reward_attr='mean_accuracy',
+            metric='mean_accuracy',
+            mode='max',
             perturbation_interval=600.0,
             hyperparam_mutations={
                 "lr": [1e-3, 5e-4, 1e-4, 5e-5, 1e-5],
@@ -52,7 +53,8 @@ The `asynchronous version of HyperBand <https://openreview.net/forum?id=S1Y7OOlR
 
     async_hb_scheduler = AsyncHyperBandScheduler(
         time_attr='training_iteration',
-        reward_attr='episode_reward_mean',
+        metric='episode_reward_mean',
+        mode='max',
         max_t=100,
         grace_period=10,
         reduction_factor=3,
diff --git a/python/ray/tune/examples/async_hyperband_example.py b/python/ray/tune/examples/async_hyperband_example.py
index 1569ced23cc9..4d85ed2ccc0c 100644
--- a/python/ray/tune/examples/async_hyperband_example.py
+++ b/python/ray/tune/examples/async_hyperband_example.py
@@ -59,7 +59,8 @@ def _restore(self, checkpoint_path):
     # which is automatically filled by Tune.
     ahb = AsyncHyperBandScheduler(
         time_attr="training_iteration",
-        reward_attr="episode_reward_mean",
+        metric="episode_reward_mean",
+        mode="max",
         grace_period=5,
         max_t=100)
 
diff --git a/python/ray/tune/examples/ax_example.py b/python/ray/tune/examples/ax_example.py
index 07bb7f79a1f3..9e82e97118ad 100644
--- a/python/ray/tune/examples/ax_example.py
+++ b/python/ray/tune/examples/ax_example.py
@@ -109,5 +109,5 @@ def easy_objective(config, reporter):
         parameter_constraints=["x1 + x2 <= 2.0"],  # Optional.
         outcome_constraints=["l2norm <= 1.25"],  # Optional.
     )
-    scheduler = AsyncHyperBandScheduler(reward_attr="hartmann6")
+    scheduler = AsyncHyperBandScheduler(metric="hartmann6", mode="max")
     run(easy_objective, name="ax", search_alg=algo, **config)
diff --git a/python/ray/tune/examples/bayesopt_example.py b/python/ray/tune/examples/bayesopt_example.py
index 3b1e8342c60a..5ba8fa8610b4 100644
--- a/python/ray/tune/examples/bayesopt_example.py
+++ b/python/ray/tune/examples/bayesopt_example.py
@@ -52,7 +52,7 @@ def easy_objective(config, reporter):
             "kappa": 2.5,
             "xi": 0.0
         })
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective,
         name="my_exp",
         search_alg=algo,
diff --git a/python/ray/tune/examples/genetic_example.py b/python/ray/tune/examples/genetic_example.py
index 12f80ebe73f9..fdcf54881d3b 100644
--- a/python/ray/tune/examples/genetic_example.py
+++ b/python/ray/tune/examples/genetic_example.py
@@ -50,7 +50,7 @@ def michalewicz_function(config, reporter):
         reward_attr="neg_mean_loss",
         max_generation=2 if args.smoke_test else 10,
         population_size=10 if args.smoke_test else 50)
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(michalewicz_function,
         name="my_exp",
         search_alg=algo,
diff --git a/python/ray/tune/examples/hyperband_example.py b/python/ray/tune/examples/hyperband_example.py
index d5ad24dbcfb7..2adf41b6ad98 100755
--- a/python/ray/tune/examples/hyperband_example.py
+++ b/python/ray/tune/examples/hyperband_example.py
@@ -58,7 +58,8 @@ def _restore(self, checkpoint_path):
     # which is automatically filled by Tune.
     hyperband = HyperBandScheduler(
         time_attr="training_iteration",
-        reward_attr="episode_reward_mean",
+        metric="episode_reward_mean",
+        mode="max",
         max_t=100)
 
     exp = Experiment(
diff --git a/python/ray/tune/examples/hyperopt_example.py b/python/ray/tune/examples/hyperopt_example.py
index 870b13d59621..26dd44109168 100644
--- a/python/ray/tune/examples/hyperopt_example.py
+++ b/python/ray/tune/examples/hyperopt_example.py
@@ -68,5 +68,5 @@ def easy_objective(config, reporter):
         max_concurrent=4,
         reward_attr="neg_mean_loss",
         points_to_evaluate=current_best_params)
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective, search_alg=algo, scheduler=scheduler, **config)
diff --git a/python/ray/tune/examples/mnist_pytorch.py b/python/ray/tune/examples/mnist_pytorch.py
index 188a8baeca36..0996b6abe413 100644
--- a/python/ray/tune/examples/mnist_pytorch.py
+++ b/python/ray/tune/examples/mnist_pytorch.py
@@ -161,7 +161,8 @@ def test():
     ray.init()
     sched = AsyncHyperBandScheduler(
         time_attr="training_iteration",
-        reward_attr="neg_mean_loss",
+        metric="neg_mean_loss",
+        mode="max",
         max_t=400,
         grace_period=20)
     tune.register_trainable(
diff --git a/python/ray/tune/examples/mnist_pytorch_trainable.py b/python/ray/tune/examples/mnist_pytorch_trainable.py
index 7163dcfd6a01..1c287221b6e2 100644
--- a/python/ray/tune/examples/mnist_pytorch_trainable.py
+++ b/python/ray/tune/examples/mnist_pytorch_trainable.py
@@ -180,7 +180,7 @@ def _restore(self, checkpoint_path):
 
     ray.init(redis_address=args.redis_address)
     sched = HyperBandScheduler(
-        time_attr="training_iteration", reward_attr="neg_mean_loss")
+        time_attr="training_iteration", metric="neg_mean_loss", mode="max")
     tune.run(
         TrainMNIST,
         scheduler=sched,
diff --git a/python/ray/tune/examples/nevergrad_example.py b/python/ray/tune/examples/nevergrad_example.py
index 3c63dea0a3bf..e17fda137ebf 100644
--- a/python/ray/tune/examples/nevergrad_example.py
+++ b/python/ray/tune/examples/nevergrad_example.py
@@ -56,7 +56,7 @@ def easy_objective(config, reporter):
         parameter_names,
         max_concurrent=4,
         reward_attr="neg_mean_loss")
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective,
         name="nevergrad",
         search_alg=algo,
diff --git a/python/ray/tune/examples/pbt_example.py b/python/ray/tune/examples/pbt_example.py
index 7c8dda4831ae..3b49a92d7c57 100755
--- a/python/ray/tune/examples/pbt_example.py
+++ b/python/ray/tune/examples/pbt_example.py
@@ -96,7 +96,8 @@ def reset_config(self, new_config):
 
     pbt = PopulationBasedTraining(
         time_attr="training_iteration",
-        reward_attr="mean_accuracy",
+        metric="mean_accuracy",
+        mode="max",
         perturbation_interval=20,
         hyperparam_mutations={
             # distribution for resampling
diff --git a/python/ray/tune/examples/pbt_ppo_example.py b/python/ray/tune/examples/pbt_ppo_example.py
index 7ccd7aa8423f..b555a68acb9a 100755
--- a/python/ray/tune/examples/pbt_ppo_example.py
+++ b/python/ray/tune/examples/pbt_ppo_example.py
@@ -30,7 +30,8 @@ def explore(config):
 
     pbt = PopulationBasedTraining(
         time_attr="time_total_s",
-        reward_attr="episode_reward_mean",
+        metric="episode_reward_mean",
+        mode="max",
         perturbation_interval=120,
         resample_probability=0.25,
         # Specifies the mutations of these hyperparams
diff --git a/python/ray/tune/examples/pbt_tune_cifar10_with_keras.py b/python/ray/tune/examples/pbt_tune_cifar10_with_keras.py
index 52327c1f1867..c1a1765bc9f4 100755
--- a/python/ray/tune/examples/pbt_tune_cifar10_with_keras.py
+++ b/python/ray/tune/examples/pbt_tune_cifar10_with_keras.py
@@ -206,7 +206,8 @@ def _stop(self):
 
     pbt = PopulationBasedTraining(
         time_attr="training_iteration",
-        reward_attr="mean_accuracy",
+        metric="mean_accuracy",
+        mode="max",
         perturbation_interval=10,
         hyperparam_mutations={
             "dropout": lambda _: np.random.uniform(0, 1),
diff --git a/python/ray/tune/examples/sigopt_example.py b/python/ray/tune/examples/sigopt_example.py
index fe2ef21f23a2..9cb85a75eb30 100644
--- a/python/ray/tune/examples/sigopt_example.py
+++ b/python/ray/tune/examples/sigopt_example.py
@@ -69,7 +69,7 @@ def easy_objective(config, reporter):
         name="SigOpt Example Experiment",
         max_concurrent=1,
         reward_attr="neg_mean_loss")
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective,
         name="my_exp",
         search_alg=algo,
diff --git a/python/ray/tune/examples/skopt_example.py b/python/ray/tune/examples/skopt_example.py
index 906746f2c03f..d34e95ef4ddd 100644
--- a/python/ray/tune/examples/skopt_example.py
+++ b/python/ray/tune/examples/skopt_example.py
@@ -51,7 +51,7 @@ def easy_objective(config, reporter):
         reward_attr="neg_mean_loss",
         points_to_evaluate=previously_run_params,
         evaluated_rewards=known_rewards)
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective,
         name="skopt_exp_with_warmstart",
         search_alg=algo,
@@ -65,7 +65,7 @@ def easy_objective(config, reporter):
         max_concurrent=4,
         reward_attr="neg_mean_loss",
         points_to_evaluate=previously_run_params)
-    scheduler = AsyncHyperBandScheduler(reward_attr="neg_mean_loss")
+    scheduler = AsyncHyperBandScheduler(metric="neg_mean_loss", mode="max")
     run(easy_objective,
         name="skopt_exp",
         search_alg=algo,
diff --git a/python/ray/tune/examples/tune_cifar10_gluon.py b/python/ray/tune/examples/tune_cifar10_gluon.py
index 63db20186b91..24ed874a2033 100644
--- a/python/ray/tune/examples/tune_cifar10_gluon.py
+++ b/python/ray/tune/examples/tune_cifar10_gluon.py
@@ -192,7 +192,8 @@ def test():
     elif args.scheduler == "asynchyperband":
         sched = AsyncHyperBandScheduler(
             time_attr="training_iteration",
-            reward_attr="neg_mean_loss",
+            metric="neg_mean_loss",
+            mode="max",
             max_t=400,
             grace_period=60)
     else:
diff --git a/python/ray/tune/examples/tune_mnist_async_hyperband.py b/python/ray/tune/examples/tune_mnist_async_hyperband.py
index 1fc92614eb62..b58ac0353ab1 100755
--- a/python/ray/tune/examples/tune_mnist_async_hyperband.py
+++ b/python/ray/tune/examples/tune_mnist_async_hyperband.py
@@ -240,7 +240,8 @@ def train(config={"activation": "relu"}, reporter=None):
         name="tune_mnist_test",
         scheduler=AsyncHyperBandScheduler(
             time_attr="timesteps_total",
-            reward_attr="mean_accuracy",
+            metric="mean_accuracy",
+            mode="max",
             max_t=600,
         ),
         **mnist_spec)
diff --git a/python/ray/tune/examples/tune_mnist_keras.py b/python/ray/tune/examples/tune_mnist_keras.py
index 676f650cca0d..5357d86af19e 100644
--- a/python/ray/tune/examples/tune_mnist_keras.py
+++ b/python/ray/tune/examples/tune_mnist_keras.py
@@ -64,7 +64,8 @@ def train_mnist(config, reporter):
     ray.init()
     sched = AsyncHyperBandScheduler(
         time_attr="timesteps_total",
-        reward_attr="mean_accuracy",
+        metric="mean_accuracy",
+        mode="max",
         max_t=400,
         grace_period=20)
 
diff --git a/python/ray/tune/examples/tune_mnist_ray_hyperband.py b/python/ray/tune/examples/tune_mnist_ray_hyperband.py
index 9401fb38cbe3..ac90d538c6b9 100755
--- a/python/ray/tune/examples/tune_mnist_ray_hyperband.py
+++ b/python/ray/tune/examples/tune_mnist_ray_hyperband.py
@@ -233,7 +233,7 @@ def _restore(self, ckpt_data):
 
     ray.init()
     hyperband = HyperBandScheduler(
-        time_attr="training_iteration", reward_attr="mean_accuracy", max_t=10)
+        time_attr="training_iteration", metric="mean_accuracy", mode="max", max_t=10)
 
     tune.run(
         TrainMNIST,
diff --git a/python/ray/tune/tests/test_experiment_analysis.py b/python/ray/tune/tests/test_experiment_analysis.py
index 96b08c0a6685..a0721abc5d29 100644
--- a/python/ray/tune/tests/test_experiment_analysis.py
+++ b/python/ray/tune/tests/test_experiment_analysis.py
@@ -36,7 +36,8 @@ def tearDown(self):
     def run_test_exp(self):
         ahb = AsyncHyperBandScheduler(
             time_attr="training_iteration",
-            reward_attr=self.metric,
+            metric=self.metric,
+            mode="max",
             grace_period=5,
             max_t=100)
 
diff --git a/python/ray/tune/tests/test_trial_scheduler.py b/python/ray/tune/tests/test_trial_scheduler.py
index 1f8ab8025f1f..553477c150b2 100644
--- a/python/ray/tune/tests/test_trial_scheduler.py
+++ b/python/ray/tune/tests/test_trial_scheduler.py
@@ -143,7 +143,8 @@ def result2(t, rew):
             grace_period=0,
             min_samples_required=1,
             time_attr="training_iteration",
-            reward_attr="neg_mean_loss")
+            metric="neg_mean_loss",
+            mode="max")
         t1 = Trial("PPO")  # mean is 450, max 900, t_max=10
         t2 = Trial("PPO")  # mean is 450, max 450, t_max=5
         for i in range(10):
@@ -502,7 +503,7 @@ def result2(t, rew):
             return dict(time_total_s=t, neg_mean_loss=rew)
 
         sched = HyperBandScheduler(
-            time_attr="time_total_s", reward_attr="neg_mean_loss")
+            time_attr="time_total_s", metric="neg_mean_loss", mode="max")
         stats = self.default_statistics()
 
         for i in range(stats["max_trials"]):
@@ -1022,7 +1023,8 @@ def result2(t, rew):
         scheduler = AsyncHyperBandScheduler(
             grace_period=1,
             time_attr="training_iteration",
-            reward_attr="neg_mean_loss",
+            metric="neg_mean_loss",
+            mode="max",
             brackets=1)
         t1 = Trial("PPO")  # mean is 450, max 900, t_max=10
         t2 = Trial("PPO")  # mean is 450, max 450, t_max=5