Learning2RL/cartpole.py at master · Phu-archive/Learning2RL · 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
123
124
125
126
127
128
129
130
131
import tensorflow as tf

import ray
import ray.tune as tune

ray.init()

import gym
import numpy as np

gamma = 0.99

class Policy(tf.keras.Model):
  def __init__(self):
    super().__init__()
    self.fc1 = tf.keras.layers.Dense(128, use_bias=True)
    self.fc2 = tf.keras.layers.Dense(10, use_bias=True)
    self.fc3 = tf.keras.layers.Dense(2, use_bias=True)
    self.fc_value = tf.keras.layers.Dense(1, use_bias=True)

  def call(self, x):
    x = tf.nn.relu(self.fc1(x))
    x_out = tf.nn.relu(self.fc2(x))
    x = self.fc3(x_out)

    val = self.fc_value(x_out)

    return x, tf.nn.softmax(x), val

def cal_discounted_reward(rewards):
    discounted_reward = []
    cumulative_sum = 0
    for i, r in enumerate(reversed(rewards)):
        cumulative_sum = (cumulative_sum + r)*gamma
        discounted_reward.append(cumulative_sum)
    return discounted_reward[::-1]

@ray.remote
def train_fuction(config, reporter):
    tf.enable_eager_execution()
    print(f"Running with config {config}")
    env = gym.make('CartPole-v1')
    obs = env.reset()
    agent = Policy()

    observation, action_taken, reward = [], [], []
    sum_reward = 0

    eps_num = 1

    optimizers = tf.train.RMSPropOptimizer(learning_rate=config["lr"])
    while True:
        _, out, _ = agent(tf.expand_dims(tf.convert_to_tensor(obs), 0))
        choice = np.random.choice(2, p=out.numpy()[0])
        observation.append(obs)

        if choice == 0:
            action_taken.append([1.0, 0.0])
        else:
            action_taken.append([0.0, 1.0])

        obs, r, done, _ = env.step(choice)
        reward.append(r)

        sum_reward += r

        grad = None
        if done:
            print(f"At Eps: {eps_num} reward is {sum_reward}")

            discounted_rewards = cal_discounted_reward(reward)
            discounted_rewards -= np.mean(discounted_rewards)
            discounted_rewards /= np.std(discounted_rewards)

            observation_numpy, action_taken_numpy, reward_numpy = np.array(observation), np.array(action_taken), np.array(reward)

            with tf.GradientTape() as tape:
                tape.watch(agent.variables)

                total_loss = 0

                for i in range(len(observation)-1):
                    fin_layer, pred_prob, value = agent(tf.convert_to_tensor([observation[i]]))
                    _, _, value_next = agent(tf.convert_to_tensor([observation[i+1]]))

                    loss = tf.nn.softmax_cross_entropy_with_logits_v2(
                        logits=fin_layer,
                        labels=tf.convert_to_tensor([action_taken_numpy[i]])
                    )


                    value_loss = tf.losses.mean_squared_error(
                        # reward[i] + 0.99 * tf.convert_to_tensor(value_next.numpy()),
                        tf.convert_to_tensor([[discounted_rewards[i]]]),
                        value
                    )

                    entropy_loss = tf.reduce_sum(pred_prob * tf.math.log(pred_prob))

                    total_loss += -loss * (discounted_rewards[i] - value) + tf.cast(value_loss, tf.float64) * 0.5 - entropy_loss * 0.01

                total_loss /= len(observation)

                # fin_layer, _, value = agent(tf.convert_to_tensor(observation_numpy))
                # total_loss =  tf.nn.softmax_cross_entropy_with_logits_v2(logits=fin_layer, labels=tf.convert_to_tensor(action_taken_numpy))
                # total_loss = tf.reduce_mean(tf.multiply(loss, discounted_rewards))
                grad = tape.gradient(total_loss, agent.variables)

            optimizers.apply_gradients(zip(grad, agent.variables))

            observation, action_taken, reward = [], [], []
            obv = env.reset()
            grad = None

            if eps_num == 800:
                break
            else:
                sum_reward = 0
            eps_num += 1

    # reporter(final_sum=sum_reward)
    print(f"Done!! {config} with {sum_reward}")
    return sum_reward

# possible_lr = [n*1e-3 for n in range(1, 10)] + [n*1e-4 for n in range(1, 10)] + [n*1e-5 for n in range(1, 10)]
# a = [train_fuction.remote({"lr": i}, print) for i in possible_lr]
# total = list(zip(ray.get(a), possible_lr))
#
# print(sorted(total, key=lambda x: x[0])[::-1][:3])

ray.get(train_fuction.remote({"lr": 1e-3}, print))