DeepRL/deep_rl/component/task.py

#######################################################################
# Copyright (C) 2017 Shangtong Zhang(zhangshangtong.cpp@gmail.com)    #
# Permission given to modify the code as long as you keep this        #
# declaration at the top                                              #
#######################################################################
from .atari_wrapper import *
import multiprocessing as mp
import sys
from .bench import Monitor
from ..utils import *
import uuid

class BaseTask:
    def set_monitor(self, env, log_dir):
        if log_dir is None:
            return env
        mkdir(log_dir)
        return Monitor(env, '%s/%s' % (log_dir, uuid.uuid4()))

    def reset(self):
        return self.env.reset()

    def step(self, action):
        return self.env.step(action)

    def seed(self, random_seed):
        return self.env.seed(random_seed)

class ClassicalControl(BaseTask):
    def __init__(self, name='CartPole-v0', max_steps=200, log_dir=None):
        BaseTask.__init__(self)
        self.name = name
        self.env = gym.make(self.name)
        self.env._max_episode_steps = max_steps
        self.action_dim = self.env.action_space.n
        self.state_dim = self.env.observation_space.shape[0]
        self.env = self.set_monitor(self.env, log_dir)

class PixelAtari(BaseTask):
    def __init__(self, name, seed=0, log_dir=None,
                 frame_skip=4, history_length=4, dataset=False):
        BaseTask.__init__(self)
        env = make_atari(name, frame_skip)
        env.seed(seed)
        if dataset:
            env = DatasetEnv(env)
            self.dataset_env = env
        env = self.set_monitor(env, log_dir)
        env = wrap_deepmind(env, history_length=history_length)
        self.env = env
        self.action_dim = self.env.action_space.n
        self.state_dim = self.env.observation_space.shape
        self.name = name

class RamAtari(BaseTask):
    def __init__(self, name, no_op, frame_skip, log_dir=None):
        BaseTask.__init__(self)
        self.name = name
        env = gym.make(name)
        assert 'NoFrameskip' in env.spec.id
        env = self.set_monitor(env, log_dir)
        env = EpisodicLifeEnv(env)
        env = NoopResetEnv(env, noop_max=no_op)
        env = SkipEnv(env, skip=frame_skip)
        if 'FIRE' in env.unwrapped.get_action_meanings():
            env = FireResetEnv(env)
        self.env = env
        self.action_dim = self.env.action_space.n
        self.state_dim = 128

class Pendulum(BaseTask):
    def __init__(self, log_dir=None):
        BaseTask.__init__(self)
        self.name = 'Pendulum-v0'
        self.env = gym.make(self.name)
        self.action_dim = self.env.action_space.shape[0]
        self.state_dim = self.env.observation_space.shape[0]
        self.env = self.set_monitor(self.env, log_dir)

    def step(self, action):
        return BaseTask.step(self, np.clip(2 * action, -2, 2))

class Box2DContinuous(BaseTask):
    def __init__(self, name, log_dir=None):
        BaseTask.__init__(self)
        self.name = name
        self.env = gym.make(self.name)
        self.action_dim = self.env.action_space.shape[0]
        self.state_dim = self.env.observation_space.shape[0]
        self.env = self.set_monitor(self.env, log_dir)

    def step(self, action):
        return BaseTask.step(self, np.clip(action, -1, 1))

class Roboschool(BaseTask):
    def __init__(self, name, log_dir=None):
        import roboschool
        BaseTask.__init__(self)
        self.name = name
        self.env = gym.make(self.name)
        self.action_dim = self.env.action_space.shape[0]
        self.state_dim = self.env.observation_space.shape[0]
        self.env = self.set_monitor(self.env, log_dir)

    def step(self, action):
        return BaseTask.step(self, np.clip(action, -1, 1))

class ProcessWrapper(mp.Process):
    STEP = 0
    RESET = 1
    EXIT = 2
    SPECS = 3
    def __init__(self, pipe, task_fn, rank, log_dir):
        mp.Process.__init__(self)
        self.pipe = pipe
        self.task_fn = task_fn
        self.log_dir = log_dir

    def run(self):
        np.random.seed()
        seed = np.random.randint(0, sys.maxsize)
        task = self.task_fn(log_dir=self.log_dir)
        task.seed(seed)
        while True:
            op, data = self.pipe.recv()
            if op == self.STEP:
                ob, reward, done, info = task.step(data)
                if done:
                    ob = task.reset()
                self.pipe.send([ob, reward, done, info])
            elif op == self.RESET:
                self.pipe.send(task.reset())
            elif op == self.EXIT:
                self.pipe.close()
                return
            elif op == self.SPECS:
                self.pipe.send([task.state_dim, task.action_dim, task.name])
            else:
                raise Exception('Unknown command')

class ParallelizedTask:
    def __init__(self, task_fn, num_workers, log_dir=None):
        self.task_fn = task_fn
        if log_dir is not None:
            mkdir(log_dir)
        self.pipes, worker_pipes = zip(*[mp.Pipe() for _ in range(num_workers)])
        args = [(wp, task_fn, rank, log_dir)
                for rank, wp in enumerate(worker_pipes)]
        self.workers = [ProcessWrapper(*arg) for arg in args]
        for p in self.workers: p.start()
        self.pipes[0].send([ProcessWrapper.SPECS, None])
        self.state_dim, self.action_dim, self.name = self.pipes[0].recv()

    def step(self, actions):
        for pipe, action in zip(self.pipes, actions):
            pipe.send((ProcessWrapper.STEP, action))
        results = [p.recv() for p in self.pipes]
        results = map(lambda x: np.stack(x), zip(*results))
        return results

    def reset(self, i=None):
        if i is None:
            for pipe in self.pipes:
                pipe.send((ProcessWrapper.RESET, None))
            results = [p.recv() for p in self.pipes]
        else:
            self.pipes[i].send((ProcessWrapper.RESET, None))
            results = self.pipes[i].recv()
        return np.stack(results)

    def close(self):
        for pipe in self.pipes:
            pipe.send((ProcessWrapper.EXIT, None))
        for p in self.workers: p.join()