policy_gradient/pendulum_a2c.py at master · ufownl/policy_gradient · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
import pickle
import argparse
import mxnet as mx
from environment import run
from distributions import Normal
from utils import AgentBase


class Actor(mx.gluon.nn.Block):
    def __init__(self, dims=128, **kwargs):
        super(Actor, self).__init__(**kwargs)
        with self.name_scope():
            self.__hidden = mx.gluon.nn.Dense(dims, activation="relu")
            self.__mu = mx.gluon.nn.Dense(1, activation="tanh", weight_initializer=mx.initializer.Uniform())
            self.__std = mx.gluon.nn.Dense(1, activation="softrelu", weight_initializer=mx.initializer.Uniform())

    def forward(self, x):
        y = self.__hidden(x)
        return Normal(self.__mu(y) * 2, mx.nd.exp(self.__std(y)))


class Critic(mx.gluon.nn.Block):
    def __init__(self, dims=128, **kwargs):
        super(Critic, self).__init__(**kwargs)
        self.__net = mx.gluon.nn.Sequential()
        with self.name_scope():
            self.__net.add(
                mx.gluon.nn.Dense(dims, activation="relu"),
                mx.gluon.nn.Dense(1, weight_initializer=mx.initializer.Uniform())
            )

    def forward(self, x):
        return self.__net(x)


class Agent(AgentBase):
    def __init__(self, gamma=0.9, entropy_weight=5e-3, ctx=mx.cpu()):
        super(Agent, self).__init__("Pendulum-v1")
        self.__actor = Actor()
        self.__actor.initialize(mx.initializer.Xavier(), ctx=ctx)
        self.__actor_trainer = mx.gluon.Trainer(self.__actor.collect_params(), "Nadam", {
            "learning_rate": 1e-4
        })
        self.__critic = Critic()
        self.__critic.initialize(mx.initializer.Xavier(), ctx=ctx)
        self.__critic_trainer = mx.gluon.Trainer(self.__critic.collect_params(), "Nadam", {
            "learning_rate": 1e-3
        })
        self.__gamma = gamma
        self.__entropy_weight = entropy_weight
        self.__context = ctx

    @property
    def test_agent(self):
        return Test(self.__actor, self.__context)

    def __call__(self):
        state, _ = yield
        while not state is None:
            s = mx.nd.array(state, ctx=self.__context).expand_dims(0)
            a = self.__actor(s).sample()
            action = a.clip(-2.0, 2.0).asnumpy()[0]
            s1, r = yield action
            g = r if s1 is None else r + self.__gamma * self.__critic(mx.nd.array(s1, ctx=self.__context).expand_dims(0))
            advantage = g - self.__critic(s)
            with mx.autograd.record():
                d = self.__actor(s)
                L = -advantage * d.log_prob(a) - self.__entropy_weight * d.entropy
                L.backward()
            self.__actor_trainer.step(1)
            with mx.autograd.record():
                L = mx.nd.smooth_l1(mx.nd.abs(g - self.__critic(s)))
                L.backward()
            self.__critic_trainer.step(1)
            state = s1


class Test(AgentBase):
    def __init__(self, actor, ctx):
        super(Test, self).__init__("Pendulum-v1", True)
        self.__actor = actor
        self.__context = ctx
        self.__demo = []

    @property
    def demo(self):
        return self.__demo

    def __call__(self):
        episode = []
        state, reward = yield
        while not state is None:
            s = mx.nd.array(state, ctx=self.__context).expand_dims(0)
            action = self.__actor(s).mean.asnumpy()[0]
            episode.append((state, reward, action))
            s1, reward = yield action
            state = s1
        episode.append((None, reward, None))
        self.__demo.append(episode)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Implementation of A2C for Pendulum-v1.")
    parser.add_argument("--episodes", help="number of training episodes (default: 500)", type=int, default=500)
    parser.add_argument("--demo", help="file path of demonstrations (default: demo.pkl)", type=str, default="demo.pkl")
    parser.add_argument("--device_id", help="select device that the model using (default: 0)", type=int, default=0)
    parser.add_argument("--gpu", help="using gpu acceleration", action="store_true")
    args = parser.parse_args()

    if args.gpu:
        agent = Agent(ctx=mx.gpu(args.device_id))
    else:
        agent = Agent(ctx=mx.cpu(args.device_id))
    print("Training...", flush=True)
    run(agent, args.episodes)
    print("Testing...", flush=True)
    test = agent.test_agent
    run(test, 5)
    print("Dumping...", flush=True)
    with open(args.demo, "wb") as f:
        pickle.dump(test.demo, f)
    print("Done!", flush=True)