Add regression testing and copy over necessary changes

.gitignore

@@ -7,3 +7,5 @@ rlkit/launchers/conf_private.py
 MANIFEST
 *.egg-info
 \.idea/
+.pytest_cache
+.coverage

README.md

@@ -214,6 +214,13 @@ This requires some knowledge of AWS and/or GCP, which is beyond the scope of
 this README.
 To learn more, more about `doodad`, [go to the repository](https://github.com/vitchyr/doodad/), which is based on [this original repository](https://github.com/justinjfu/doodad/).
 
+### Testing
+A set of basic regression tests can be run with:
+```
+nose2 -v -B -s tests/regression/basic
+```
+Other directories in `tests/regression` have project-specific tests.
+
 # Requests for pull-requests
  - Implement policy-gradient algorithms.
  - Implement model-based algorithms.

examples/awac/hand/awac1.py

@@ -3,19 +3,22 @@
 
 import rlkit.util.hyperparameter as hyp
 from rlkit.launchers.launcher_util import run_experiment
+from rlkit.launchers.arglauncher import run_variants
 
 from rlkit.torch.sac.policies import GaussianPolicy
 from rlkit.torch.networks import Clamp
 
-if __name__ == "__main__":
+def main():
     variant = dict(
-        num_epochs=501,
-        num_eval_steps_per_epoch=1000,
-        num_trains_per_train_loop=1000,
-        num_expl_steps_per_train_loop=1000,
-        min_num_steps_before_training=1000,
+        algo_kwargs=dict(
+            num_epochs=501,
+            num_eval_steps_per_epoch=1000,
+            num_trains_per_train_loop=1000,
+            num_expl_steps_per_train_loop=1000,
+            min_num_steps_before_training=1000,
+            batch_size=1024,
+        ),
         max_path_length=1000,
-        batch_size=1024,
         algorithm="AWAC",
         replay_buffer_size=int(1E6),
 
@@ -99,35 +102,11 @@
         search_space, default_parameters=variant,
     )
 
-    # n_seeds = 1
-    # mode = 'local'
-    # exp_prefix = 'dev-{}'.format(
-    #     __file__.replace('/', '-').replace('_', '-').split('.')[0]
-    # )
-
-    n_seeds = 3
-    mode = 'ec2'
-    exp_prefix = 'hand-awac1'
+    variants = []
+    for variant in sweeper.iterate_hyperparameters():
+        variants.append(variant)
 
-    for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
-        for _ in range(n_seeds):
-            run_experiment(
-                experiment,
-                exp_prefix=exp_prefix,
-                mode=mode,
-                variant=variant,
-                use_gpu=True,
-                snapshot_gap=200,
-                snapshot_mode='gap_and_last',
-                num_exps_per_instance=3,
-                gcp_kwargs=dict(
-                    zone='us-west1-b',
-                ),
+    run_variants(experiment, variants, process_args)
 
-            )
-
-    # variants = []
-    # for variant in sweeper.iterate_hyperparameters():
-    #     variants.append(variant)
-
-    # run_variants(experiment, variants, process_args)
+if __name__ == "__main__":
+    main()

examples/awac/mujoco/awac1.py

@@ -3,19 +3,22 @@
 
 import rlkit.util.hyperparameter as hyp
 from rlkit.launchers.launcher_util import run_experiment
+from rlkit.launchers.arglauncher import run_variants
 
 from rlkit.torch.sac.policies import GaussianPolicy
 from rlkit.torch.networks import Clamp
 
-if __name__ == "__main__":
+def main():
     variant = dict(
-        num_epochs=501,
-        num_eval_steps_per_epoch=1000,
-        num_trains_per_train_loop=1000,
-        num_expl_steps_per_train_loop=1000,
-        min_num_steps_before_training=1000,
+        algo_kwargs=dict(
+            num_epochs=501,
+            num_eval_steps_per_epoch=1000,
+            num_trains_per_train_loop=1000,
+            num_expl_steps_per_train_loop=1000,
+            min_num_steps_before_training=1000,
+            batch_size=1024,
+        ),
         max_path_length=1000,
-        batch_size=1024,
         replay_buffer_size=int(1E6),
         layer_size=256,
         num_layers=2,
@@ -87,35 +90,11 @@
         search_space, default_parameters=variant,
     )
 
-    n_seeds = 1
-    mode = 'local'
-    exp_prefix = 'dev-{}'.format(
-        __file__.replace('/', '-').replace('_', '-').split('.')[0]
-    )
-
-    # n_seeds = 3
-    # mode = 'gcp'
-    # exp_prefix = 'skew-fit-pickup-reference-post-refactor'
+    variants = []
+    for variant in sweeper.iterate_hyperparameters():
+        variants.append(variant)
 
-    for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
-        for _ in range(n_seeds):
-            run_experiment(
-                experiment,
-                exp_prefix=exp_prefix,
-                mode=mode,
-                variant=variant,
-                use_gpu=use_gpu,
-                snapshot_gap=200,
-                snapshot_mode='gap_and_last',
-                num_exps_per_instance=3,
-                gcp_kwargs=dict(
-                    zone='us-west1-b',
-                ),
+    run_variants(experiment, variants, process_args)
 
-            )
-
-    # variants = []
-    # for variant in sweeper.iterate_hyperparameters():
-    #     variants.append(variant)
-
-    # run_variants(experiment, variants, process_args)
+if __name__ == "__main__":
+    main()

examples/ddpg.py

@@ -18,13 +18,24 @@
 import rlkit.torch.pytorch_util as ptu
 from rlkit.torch.torch_rl_algorithm import TorchBatchRLAlgorithm
 
+import random
+import torch
+import numpy as np
 
 def experiment(variant):
     eval_env = NormalizedBoxEnv(HalfCheetahEnv())
     expl_env = NormalizedBoxEnv(HalfCheetahEnv())
     # Or for a specific version:
     # import gym
     # env = NormalizedBoxEnv(gym.make('HalfCheetah-v1'))
+
+    seed = variant["seed"]
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    eval_env.seed(seed)
+    expl_env.seed(seed)
+
     obs_dim = eval_env.observation_space.low.size
     action_dim = eval_env.action_space.low.size
     qf = ConcatMlp(
@@ -65,34 +76,36 @@ def experiment(variant):
     algorithm.to(ptu.device)
     algorithm.train()
 
+variant = dict(
+    algorithm_kwargs=dict(
+        num_epochs=1000,
+        num_eval_steps_per_epoch=1000,
+        num_trains_per_train_loop=1000,
+        num_expl_steps_per_train_loop=1000,
+        min_num_steps_before_training=10000,
+        max_path_length=1000,
+        batch_size=128,
+    ),
+    trainer_kwargs=dict(
+        use_soft_update=True,
+        tau=1e-2,
+        discount=0.99,
+        qf_learning_rate=1e-3,
+        policy_learning_rate=1e-4,
+    ),
+    qf_kwargs=dict(
+        hidden_sizes=[400, 300],
+    ),
+    policy_kwargs=dict(
+        hidden_sizes=[400, 300],
+    ),
+    replay_buffer_size=int(1E6),
+    seed=random.randint(0, 100000),
+)
 
-if __name__ == "__main__":
-    # noinspection PyTypeChecker
-    variant = dict(
-        algorithm_kwargs=dict(
-            num_epochs=1000,
-            num_eval_steps_per_epoch=1000,
-            num_trains_per_train_loop=1000,
-            num_expl_steps_per_train_loop=1000,
-            min_num_steps_before_training=10000,
-            max_path_length=1000,
-            batch_size=128,
-        ),
-        trainer_kwargs=dict(
-            use_soft_update=True,
-            tau=1e-2,
-            discount=0.99,
-            qf_learning_rate=1e-3,
-            policy_learning_rate=1e-4,
-        ),
-        qf_kwargs=dict(
-            hidden_sizes=[400, 300],
-        ),
-        policy_kwargs=dict(
-            hidden_sizes=[400, 300],
-        ),
-        replay_buffer_size=int(1E6),
-    )
-    # ptu.set_gpu_mode(True)  # optionally set the GPU (default=False)
+def main():
     setup_logger('name-of-experiment', variant=variant)
     experiment(variant)
+
+if __name__ == "__main__":
+    main()

examples/dqn_and_double_dqn.py

@@ -17,13 +17,23 @@
 from rlkit.samplers.data_collector import MdpPathCollector
 from rlkit.torch.torch_rl_algorithm import TorchBatchRLAlgorithm
 
+import random
+import torch
+import numpy as np
 
 def experiment(variant):
     expl_env = gym.make('CartPole-v0').env
     eval_env = gym.make('CartPole-v0').env
     obs_dim = expl_env.observation_space.low.size
     action_dim = eval_env.action_space.n
 
+    seed = variant["seed"]
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    eval_env.seed(seed)
+    expl_env.seed(seed)
+
     qf = Mlp(
         hidden_sizes=[32, 32],
         input_size=obs_dim,
@@ -70,28 +80,31 @@ def experiment(variant):
     algorithm.to(ptu.device)
     algorithm.train()
 
+variant = dict(
+    algorithm="DQN",
+    version="normal",
+    layer_size=256,
+    replay_buffer_size=int(1E6),
+    algorithm_kwargs=dict(
+        num_epochs=3000,
+        num_eval_steps_per_epoch=5000,
+        num_trains_per_train_loop=1000,
+        num_expl_steps_per_train_loop=1000,
+        min_num_steps_before_training=1000,
+        max_path_length=1000,
+        batch_size=256,
+    ),
+    trainer_kwargs=dict(
+        discount=0.99,
+        learning_rate=3E-4,
+    ),
+    seed=random.randint(0, 100000),
+)
 
-if __name__ == "__main__":
-    # noinspection PyTypeChecker
-    variant = dict(
-        algorithm="DQN",
-        version="normal",
-        layer_size=256,
-        replay_buffer_size=int(1E6),
-        algorithm_kwargs=dict(
-            num_epochs=3000,
-            num_eval_steps_per_epoch=5000,
-            num_trains_per_train_loop=1000,
-            num_expl_steps_per_train_loop=1000,
-            min_num_steps_before_training=1000,
-            max_path_length=1000,
-            batch_size=256,
-        ),
-        trainer_kwargs=dict(
-            discount=0.99,
-            learning_rate=3E-4,
-        ),
-    )
+def main():
     setup_logger('dqn-CartPole', variant=variant)
     # ptu.set_gpu_mode(True)  # optionally set the GPU (default=False)
     experiment(variant)
+
+if __name__ == "__main__":
+    main()