Move garage package to src/ (rlworkgroup#665)

This simplifies packaging for pip, and complies with Python packaging
best practices.

MANIFEST.in

@@ -0,0 +1,27 @@
+# project metadata
+#
+# NOTE: README.md and VERSION are required to run setup.py. Failure to include
+# them will create a broken PyPI distribution.
+include README.md
+include VERSION
+include LICENSE
+include CONTRIBUTING.md
+include CHANGELOG.md
+
+# tests
+graft tests
+include setup.cfg
+
+# documentation
+graft docs
+prune docs/_build
+
+# examples, scripts, etc.
+include Makefile
+graft docker
+graft examples
+graft scripts
+
+# ignored files
+global-exclude *.py[co]
+global-exclude .DS_Store

docker/Dockerfile.base

@@ -87,10 +87,12 @@ RUN conda update -q -y conda
 # - README.md
 # - VERSION
 # - scripts/garage
+# - src/garage/__init__.py
 # - setup.py
 COPY README.md /root/code/garage/README.md
 COPY VERSION /root/code/garage/VERSION
 COPY scripts/garage /root/code/garage/scripts/garage
+COPY src/garage/__init__.py /root/code/garage/src/garage/__init__.py
 COPY setup.py /root/code/garage/setup.py
 COPY environment.yml /root/code/garage/environment.yml
 

examples/np/cem_cartpole.py

@@ -19,10 +19,10 @@
 def run_task(*_):
     """Train CEM with Cartpole-v1 environment."""
     with LocalRunner() as runner:
-        env = TfEnv(env_name="CartPole-v1")
+        env = TfEnv(env_name='CartPole-v1')
 
         policy = CategoricalMLPPolicy(
-            name="policy", env_spec=env.spec, hidden_sizes=(32, 32))
+            name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
 
         baseline = LinearFeatureBaseline(env_spec=env.spec)
 
@@ -45,6 +45,6 @@ def run_task(*_):
 
 run_experiment(
     run_task,
-    snapshot_mode="last",
+    snapshot_mode='last',
     seed=1,
 )

examples/np/cma_es_cartpole.py

@@ -19,10 +19,10 @@
 def run_task(*_):
     """Train CMA_ES with Cartpole-v1 environment."""
     with LocalRunner() as runner:
-        env = TfEnv(env_name="CartPole-v1")
+        env = TfEnv(env_name='CartPole-v1')
 
         policy = CategoricalMLPPolicy(
-            name="policy", env_spec=env.spec, hidden_sizes=(32, 32))
+            name='policy', env_spec=env.spec, hidden_sizes=(32, 32))
 
         baseline = LinearFeatureBaseline(env_spec=env.spec)
 
@@ -44,6 +44,6 @@ def run_task(*_):
 
 run_experiment(
     run_task,
-    snapshot_mode="last",
+    snapshot_mode='last',
     seed=1,
 )

examples/tf/dqn_cartpole.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 An example to train a task with DQN algorithm.
 

examples/tf/dqn_pong.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 This is an example to train a task with DQN algorithm in pixel environment.
 

examples/tf/reps_gym_cartpole.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 This is an example to train a task with REPS algorithm.
 

scripts/travisci/check_flake8.sh

@@ -141,7 +141,7 @@ check_flake8() {
   flake8 --isolated \
          --import-order-style=google \
          --application-import-names="${garage_packages}" \
-         --per-file-ignores="./garage/misc/krylov.py:N802,N803,N806" \
+         --per-file-ignores="./src/garage/misc/krylov.py:N802,N803,N806" \
          "$@"
   status="$((${status} | ${?}))"
 }

setup.cfg

@@ -10,7 +10,7 @@ ignore-names = setUp, tearDown, setUpClass,tearDownClass, setUpModule, tearDownM
 import-order-style = google
 application-import-names = tests, sandbox, garage, examples
 per-file-ignores =
-    ./garage/misc/krylov.py:N802,N803,N806
+    ./src/garage/misc/krylov.py:N802,N803,N806
 
 [pylint]
 #[MESSAGES CONTROL]
@@ -87,4 +87,4 @@ allow_multiline_lambdas = true
 
 [coverage:run]
 branch = true
-source = ./garage
+source = src

setup.py

@@ -60,14 +60,13 @@
     version = v.read().strip()
 
 setup(
-    name='rlgarage',
+    name='garage',
     version=version,
     author='Reinforcement Learning Working Group',
     description='A framework for reproducible reinforcement learning research',
     url='https://github.com/rlworkgroup/garage',
-    packages=[
-        package for package in find_packages() if package.startswith('garage')
-    ],
+    packages=find_packages(where='src'),
+    package_dir={'': 'src'},
     scripts=['scripts/garage'],
     install_requires=required,
     extras_require=extras,

garage/config.py → src/garage/config.py

@@ -45,8 +45,8 @@
 GARAGE_AWS_KEY_NAME = os.environ.get('GARAGE_AWS_KEY_NAME', None)
 GARAGE_AWS_SPOT = bool(os.environ.get('GARAGE_AWS_SPOT', True))
 GARAGE_AWS_SPOT_PRICE = os.environ.get('GARAGE_AWS_SPOT_PRICE', '1.0')
-GARAGE_AWS_ACCESS_KEY = os.environ.get("GARAGE_AWS_ACCESS_KEY", None)
-GARAGE_AWS_ACCESS_SECRET = os.environ.get("GARAGE_AWS_ACCESS_SECRET", None)
+GARAGE_AWS_ACCESS_KEY = os.environ.get('GARAGE_AWS_ACCESS_KEY', None)
+GARAGE_AWS_ACCESS_SECRET = os.environ.get('GARAGE_AWS_ACCESS_SECRET', None)
 GARAGE_AWS_IAM_INSTANCE_PROFILE_NAME = os.environ.get(
     'GARAGE_AWS_IAM_INSTANCE_PROFILE_NAME', 'garage')
 GARAGE_AWS_SECURITY_GROUPS = eval(

src/garage/contrib/alexbeloi/examples/trpois_inverted_pendulum.py

@@ -0,0 +1,50 @@
+"""
+Example using TRPO with ISSampler.
+
+Iterations alternate between live and importance sampled iterations.
+"""
+import gym
+
+from garage.contrib.alexbeloi.is_sampler import ISSampler
+from garage.envs import normalize
+from garage.experiment import LocalRunner, run_experiment
+from garage.np.baselines import LinearFeatureBaseline
+from garage.tf.algos import TRPO
+from garage.tf.envs import TfEnv
+from garage.tf.policies import GaussianMLPPolicy
+
+
+def run_task(*_):
+    """Run the job."""
+    with LocalRunner() as runner:
+        env = TfEnv(normalize(gym.make('InvertedPendulum-v2')))
+
+        policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(32, 32))
+
+        baseline = LinearFeatureBaseline(env_spec=env.spec)
+
+        optimizer_args = dict(
+            # debug_nan=True,
+            # reg_coeff=0.1,
+            # cg_iters=2
+        )
+
+        algo = TRPO(
+            env_spec=env.spec,
+            policy=policy,
+            baseline=baseline,
+            max_path_length=100,
+            discount=0.99,
+            max_kl_step=0.01,
+            optimizer_args=optimizer_args)
+
+        runner.setup(
+            algo, env, sampler_cls=ISSampler, sampler_args=dict(n_backtrack=1))
+        runner.train(n_epochs=200, batch_size=4000)
+
+
+run_experiment(
+    run_task,
+    snapshot_mode='last',
+    seed=1,
+)

src/garage/contrib/alexbeloi/examples/vpgis_inverted_pendulum.py

@@ -0,0 +1,44 @@
+"""
+Example using VPG with ISSampler.
+
+Iterations alternate between live and importance sampled iterations.
+"""
+import gym
+
+from garage.contrib.alexbeloi.is_sampler import ISSampler
+from garage.envs import normalize
+from garage.experiment import LocalRunner, run_experiment
+from garage.np.baselines import LinearFeatureBaseline
+from garage.tf.algos import VPG
+from garage.tf.envs import TfEnv
+from garage.tf.policies import GaussianMLPPolicy
+
+
+def run_task(*_):
+    """Run the job."""
+    with LocalRunner() as runner:
+        env = TfEnv(normalize(gym.make('InvertedPendulum-v2')))
+
+        policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(32, 32))
+
+        baseline = LinearFeatureBaseline(env_spec=env.spec)
+
+        algo = VPG(
+            env_spec=env.spec,
+            policy=policy,
+            baseline=baseline,
+            max_path_length=100,
+            discount=0.99,
+            max_kl_step=0.01,
+        )
+
+        runner.setup(
+            algo, env, sampler_cls=ISSampler, sampler_args=dict(n_backtrack=1))
+        runner.train(n_epochs=40, batch_size=4000)
+
+
+run_experiment(
+    run_task,
+    snapshot_mode='last',
+    seed=1,
+)

garage/core/__init__.py → src/garage/core/__init__.py

@@ -1,4 +1,4 @@
 from garage.core.serializable import Serializable
 from garage.core.parameterized import Parameterized  # noqa: I100
 
-__all__ = ["Serializable", "Parameterized"]
+__all__ = ['Serializable', 'Parameterized']

garage/core/parameterized.py → src/garage/core/parameterized.py

@@ -60,27 +60,27 @@ def get_param_values(self, **tags):
         ])
 
     def set_param_values(self, flattened_params, **tags):
-        debug = tags.pop("debug", False)
+        debug = tags.pop('debug', False)
         param_values = unflatten_tensors(flattened_params,
                                          self.get_param_shapes(**tags))
         for param, dtype, value in zip(
                 self.get_params(**tags), self.get_param_dtypes(**tags),
                 param_values):
             param.set_value(value.astype(dtype))
             if debug:
-                print("setting value of %s" % param.name)
+                print('setting value of %s' % param.name)
 
     def flat_to_params(self, flattened_params, **tags):
         return unflatten_tensors(flattened_params,
                                  self.get_param_shapes(**tags))
 
     def __getstate__(self):
         d = Serializable.__getstate__(self)
-        d["params"] = self.get_param_values()
+        d['params'] = self.get_param_values()
         return d
 
     def __setstate__(self, d):
         Serializable.__setstate__(self, d)
         global load_params
         if load_params:
-            self.set_param_values(d["params"])
+            self.set_param_values(d['params'])

garage/core/serializable.py → src/garage/core/serializable.py

@@ -8,7 +8,7 @@ def __init__(self, *args, **kwargs):
         self.__kwargs = kwargs
 
     def quick_init(self, locals_):
-        if getattr(self, "_serializable_initialized", False):
+        if getattr(self, '_serializable_initialized', False):
             return
         if sys.version_info >= (3, 0):
             spec = inspect.getfullargspec(self.__init__)
@@ -30,13 +30,13 @@ def quick_init(self, locals_):
         in_order_args = [locals_[arg] for arg in spec.args][1:]
         self.__args = tuple(in_order_args) + varargs
         self.__kwargs = kwargs
-        setattr(self, "_serializable_initialized", True)
+        setattr(self, '_serializable_initialized', True)
 
     def __getstate__(self):
-        return {"__args": self.__args, "__kwargs": self.__kwargs}
+        return {'__args': self.__args, '__kwargs': self.__kwargs}
 
     def __setstate__(self, d):
-        out = type(self)(*d["__args"], **d["__kwargs"])
+        out = type(self)(*d['__args'], **d['__kwargs'])
         self.__dict__.update(out.__dict__)
 
     @classmethod
@@ -52,12 +52,12 @@ def clone(cls, obj, **kwargs):
             spec = inspect.getargspec(obj.__init__)
         in_order_args = spec.args[1:]
 
-        d["__args"] = list(d["__args"])
+        d['__args'] = list(d['__args'])
         for kw, val in kwargs.items():
             if kw in in_order_args:
-                d["__args"][in_order_args.index(kw)] = val
+                d['__args'][in_order_args.index(kw)] = val
             else:
-                d["__kwargs"][kw] = val
+                d['__kwargs'][kw] = val
 
         out = type(obj).__new__(type(obj))
         out.__setstate__(d)