ppo and trpo

Author: John Schulman

.gitignore

@@ -7,12 +7,11 @@
 # Setuptools distribution and build folders.
 /dist/
 /build
+keys/
 
 # Virtualenv
 /env
 
-# Python egg metadata, regenerated from source files by setuptools.
-/*.egg-info
 
 *.sublime-project
 *.sublime-workspace
@@ -26,4 +25,8 @@ ghostdriver.log
 
 htmlcov
 
-junk
+junk
+src
+
+*.egg-info
+.cache

README.md

@@ -1,8 +1,8 @@
 <img src="data/logo.jpg" width=25% align="right" />
 
-# Baselines
+# BASELINES
 
-We're releasing OpenAI Baselines, a set of high-quality implementations of reinforcement learning algorithms. To start, we're making available an open source version of Deep Q-Learning and three of its variants. 
+We're releasing OpenAI Baselines, a set of high-quality implementations of reinforcement learning algorithms.
 
 These algorithms will make it easier for the research community to replicate, refine, and identify new ideas, and will create good baselines to build research on top of. Our DQN implementation and its variants are roughly on par with the scores in published papers. We expect they will be used as a base around which new ideas can be added, and as a tool for comparing a new approach against existing ones. 
 
@@ -12,56 +12,6 @@ You can install it by typing:
 pip install baselines
 ```
 
-
-## If you are curious.
-
-##### Train a Cartpole agent and watch it play once it converges!
-
-Here's a list of commands to run to quickly get a working example:
-
-<img src="data/cartpole.gif" width="25%" />
-
-
-```bash
-# Train model and save the results to cartpole_model.pkl
-python -m baselines.deepq.experiments.train_cartpole
-# Load the model saved in cartpole_model.pkl and visualize the learned policy
-python -m baselines.deepq.experiments.enjoy_cartpole
-```
-
-
-Be sure to check out the source code of [both](baselines/deepq/experiments/train_cartpole.py) [files](baselines/deepq/experiments/enjoy_cartpole.py)!
-
-## If you wish to apply DQN to solve a problem.
-
-Check out our simple agent trained with one stop shop `deepq.learn` function. 
-
-- `baselines/deepq/experiments/train_cartpole.py` - train a Cartpole agent.
-- `baselines/deepq/experiments/train_pong.py` - train a Pong agent using convolutional neural networks.
-
-In particular notice that once `deepq.learn` finishes training it returns `act` function which can be used to select actions in the environment. Once trained you can easily save it and load at later time. For both of the files listed above there are complimentary files `enjoy_cartpole.py` and `enjoy_pong.py` respectively, that load and visualize the learned policy.
-
-## If you wish to experiment with the algorithm
-
-##### Check out the examples
-
-
-- `baselines/deepq/experiments/custom_cartpole.py` - Cartpole training with more fine grained control over the internals of DQN algorithm.
-- `baselines/deepq/experiments/atari/train.py` - more robust setup for training at scale.
-
-
-##### Download a pretrained Atari agent
-
-For some research projects it is sometimes useful to have an already trained agent handy. There's a variety of models to choose from. You can list them all by running:
-
-```bash
-python -m baselines.deepq.experiments.atari.download_model
-```
-
-Once you pick a model, you can download it and visualize the learned policy. Be sure to pass `--dueling` flag to visualization script when using dueling models.
-
-```bash
-python -m baselines.deepq.experiments.atari.download_model --blob model-atari-duel-pong-1 --model-dir /tmp/models
-python -m baselines.deepq.experiments.atari.enjoy --model-dir /tmp/models/model-atari-duel-pong-1 --env Pong --dueling
-
-```
+- [DQN](baselines/deepq)
+- [PPO](baselines/pposgd)
+- [TRPO](baselines/trpo_mpi)

baselines/bench/__init__.py

@@ -0,0 +1,3 @@
+from baselines.bench.benchmarks import *
+from baselines.bench.monitor import *
+

baselines/bench/benchmarks.py

@@ -0,0 +1,93 @@
+_atari7 = ['BeamRider', 'Breakout', 'Enduro', 'Pong', 'Qbert', 'Seaquest', 'SpaceInvaders']
+_atariexpl7 = ['Freeway', 'Gravitar', 'MontezumaRevenge', 'Pitfall', 'PrivateEye', 'Solaris', 'Venture']
+
+_BENCHMARKS = []
+
+def register_benchmark(benchmark):
+    for b in _BENCHMARKS:
+        if b['name'] == benchmark['name']:
+            raise ValueError('Benchmark with name %s already registered!'%b['name'])
+    _BENCHMARKS.append(benchmark)
+
+def list_benchmarks():
+    return [b['name'] for b in _BENCHMARKS]
+
+def get_benchmark(benchmark_name):
+    for b in _BENCHMARKS:
+        if b['name'] == benchmark_name:
+            return b
+    raise ValueError('%s not found! Known benchmarks: %s' % (benchmark_name, list_benchmarks()))
+
+def get_task(benchmark, env_id):
+    """Get a task by env_id. Return None if the benchmark doesn't have the env"""
+    return next(filter(lambda task: task['env_id'] == env_id, benchmark['tasks']), None)
+
+_ATARI_SUFFIX = 'NoFrameskip-v4'
+
+register_benchmark({
+    'name' : 'Atari200M',
+    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 200M frames',
+    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(200e6)} for _game in _atari7]
+})
+
+register_benchmark({
+    'name' : 'Atari40M',
+    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 40M frames',
+    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(40e6)} for _game in _atari7]
+})
+
+register_benchmark({
+    'name' : 'Atari1Hr',
+    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 1 hour of walltime',
+    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_seconds' : 60*60} for _game in _atari7]
+})
+
+register_benchmark({
+    'name' : 'AtariExploration40M',
+    'description' :'7 Atari games emphasizing exploration, with pixel observations, 40M frames',
+    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(40e6)} for _game in _atariexpl7]
+})
+
+
+_mujocosmall = [
+    'InvertedDoublePendulum-v1', 'InvertedPendulum-v1',
+    'HalfCheetah-v1', 'Hopper-v1', 'Walker2d-v1',
+    'Reacher-v1', 'Swimmer-v1']
+
+register_benchmark({
+    'name' : 'Mujoco1M',
+    'description' : 'Some small 2D MuJoCo tasks, run for 1M timesteps',
+    'tasks' : [{'env_id' : _envid, 'trials' : 3, 'num_timesteps' : int(1e6)} for _envid in _mujocosmall]
+})
+
+_roboschool_mujoco = [
+    'RoboschoolInvertedDoublePendulum-v0', 'RoboschoolInvertedPendulum-v0',      # cartpole
+    'RoboschoolHalfCheetah-v0', 'RoboschoolHopper-v0', 'RoboschoolWalker2d-v0',  # forward walkers
+    'RoboschoolReacher-v0'
+    ]
+
+register_benchmark({
+    'name' : 'RoboschoolMujoco2M',
+    'description' : 'Same small 2D tasks, still improving up to 2M',
+    'tasks' : [{'env_id' : _envid, 'trials' : 3, 'num_timesteps' : int(2e6)} for _envid in _roboschool_mujoco]
+})
+
+
+_atari50 =  [ # actually 49
+            'Alien', 'Amidar', 'Assault', 'Asterix', 'Asteroids', 
+            'Atlantis', 'BankHeist', 'BattleZone', 'BeamRider',  'Bowling', 
+            'Boxing', 'Breakout', 'Centipede', 'ChopperCommand', 'CrazyClimber', 
+            'DemonAttack', 'DoubleDunk',  'Enduro', 'FishingDerby', 'Freeway', 
+            'Frostbite', 'Gopher', 'Gravitar', 'IceHockey', 'Jamesbond',  
+            'Kangaroo', 'Krull', 'KungFuMaster', 'MontezumaRevenge', 'MsPacman', 
+            'NameThisGame', 'Pitfall', 'Pong', 'PrivateEye', 'Qbert', 
+            'Riverraid', 'RoadRunner', 'Robotank', 'Seaquest', 'SpaceInvaders', 
+            'StarGunner', 'Tennis', 'TimePilot', 'Tutankham', 'UpNDown', 
+            'Venture', 'VideoPinball', 'WizardOfWor', 'Zaxxon', 
+]
+
+register_benchmark({
+    'name' : 'Atari50_40M',
+    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 40M frames',
+    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 3, 'num_timesteps' : int(40e6)} for _game in _atari50]
+})

baselines/bench/monitor.py

@@ -0,0 +1,146 @@
+__all__ = ['Monitor', 'get_monitor_files', 'load_results']
+
+import gym
+from gym.core import Wrapper
+from os import path
+import time
+from glob import glob
+
+try:
+    import ujson as json # Not necessary for monitor writing, but very useful for monitor loading
+except ImportError:
+    import json
+
+class Monitor(Wrapper):
+    EXT = "monitor.json"
+    f = None
+
+    def __init__(self, env, filename, allow_early_resets=False):
+        Wrapper.__init__(self, env=env)
+        self.tstart = time.time()
+        if filename is None:
+            self.f = None
+            self.logger = None
+        else:
+            if not filename.endswith(Monitor.EXT):
+                filename = filename + "." + Monitor.EXT
+            self.f = open(filename, "wt")
+            self.logger = JSONLogger(self.f)
+            self.logger.writekvs({"t_start": self.tstart, "gym_version": gym.__version__,
+                "env_id": env.spec.id if env.spec else 'Unknown'})
+        self.allow_early_resets = allow_early_resets
+        self.rewards = None
+        self.needs_reset = True
+        self.episode_rewards = []
+        self.episode_lengths = []
+        self.total_steps = 0
+        self.current_metadata = {} # extra info that gets injected into each log entry
+        # Useful for metalearning where we're modifying the environment externally
+        # But want our logs to know about these modifications
+
+    def __getstate__(self): # XXX
+        d = self.__dict__.copy()
+        if self.f:
+            del d['f'], d['logger']
+            d['_filename'] = self.f.name
+            d['_num_episodes'] = len(self.episode_rewards)
+        else:
+            d['_filename'] = None
+        return d
+    def __setstate__(self, d):
+        filename = d.pop('_filename')
+        self.__dict__ = d
+        if filename is not None:
+            nlines = d.pop('_num_episodes') + 1
+            self.f = open(filename, "r+t")
+            for _ in range(nlines):
+                self.f.readline()
+            self.f.truncate()        
+            self.logger = JSONLogger(self.f)
+
+
+    def reset(self):
+        if not self.allow_early_resets and not self.needs_reset:
+            raise RuntimeError("Tried to reset an environment before done. If you want to allow early resets, wrap your env with Monitor(env, path, allow_early_resets=True)")
+        self.rewards = []
+        self.needs_reset = False
+        return self.env.reset()
+
+    def step(self, action):
+        if self.needs_reset:
+            raise RuntimeError("Tried to step environment that needs reset")
+        ob, rew, done, info = self.env.step(action)
+        self.rewards.append(rew)
+        if done:
+            self.needs_reset = True
+            eprew = sum(self.rewards)
+            eplen = len(self.rewards)
+            epinfo = {"r": eprew, "l": eplen, "t": round(time.time() - self.tstart, 6)}
+            epinfo.update(self.current_metadata)
+            if self.logger:
+                self.logger.writekvs(epinfo)
+            self.episode_rewards.append(eprew)
+            self.episode_lengths.append(eplen)
+            info['episode'] = epinfo
+        self.total_steps += 1
+        return (ob, rew, done, info)
+
+    def close(self):
+        if self.f is not None:
+            self.f.close()
+
+    def get_total_steps(self):
+        return self.total_steps
+
+    def get_episode_rewards(self):
+        return self.episode_rewards
+
+    def get_episode_lengths(self):
+        return self.episode_lengths
+
+class JSONLogger(object):
+    def __init__(self, file):
+        self.file = file
+
+    def writekvs(self, kvs):
+        for k,v in kvs.items():
+            if hasattr(v, 'dtype'):
+                v = v.tolist()
+                kvs[k] = float(v)
+        self.file.write(json.dumps(kvs) + '\n')
+        self.file.flush()
+
+
+class LoadMonitorResultsError(Exception):
+    pass
+
+def get_monitor_files(dir):
+    return glob(path.join(dir, "*" + Monitor.EXT))
+
+def load_results(dir):
+    fnames = get_monitor_files(dir)
+    if not fnames:
+        raise LoadMonitorResultsError("no monitor files of the form *%s found in %s" % (Monitor.EXT, dir))
+    episodes = []
+    headers = []
+    for fname in fnames:
+        with open(fname, 'rt') as fh:
+            lines = fh.readlines()
+        header = json.loads(lines[0])
+        headers.append(header)
+        for line in lines[1:]:
+            episode = json.loads(line)
+            episode['abstime'] = header['t_start'] + episode['t']
+            del episode['t']
+            episodes.append(episode)
+    header0 = headers[0]
+    for header in headers[1:]:
+        assert header['env_id'] == header0['env_id'], "mixing data from two envs"
+    episodes = sorted(episodes, key=lambda e: e['abstime'])
+    return {
+        'env_info': {'env_id': header0['env_id'], 'gym_version': header0['gym_version']},
+        'episode_end_times': [e['abstime'] for e in episodes],
+        'episode_lengths': [e['l'] for e in episodes],
+        'episode_rewards': [e['r'] for e in episodes],
+        'initial_reset_time': min([min(header['t_start'] for header in headers)])
+    }

baselines/common/__init__.py

@@ -1,4 +1,4 @@
-
-
-
+from baselines.common.console_util import *
+from baselines.common.dataset import Dataset
+from baselines.common.math_util import *
 from baselines.common.misc_util import *