advantage_sum_mbrl/main.py at main · re0078/advantage_sum_mbrl · 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
132
133
134
135
136
137
138
139
140
141
## script Value Summation Model-Based Reinforcement Learning
# Author   : Mehran Raisi
# Date     : 15 September 2022

import os
import argparse
import gym
import numpy as np
import tensorflow as tf
from sac import Agent
import logging
import tensorflow.keras as keras
import gc

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--env_id', type=str, help='rl environment', required=True)
    parser.add_argument('--instance_number', type=int, help='instance of the environment', default=1)
    parser.add_argument('--load_models', type=bool, help='loading trained environment', default=False)
    parser.add_argument('--load_epoch', type=int, help='loading trained environment epoch', default=None)
    parser.add_argument('--scoring_method', type=str,
                        help="scoring function approach (either 'value' or 'advantage' )", default='advantage')
    # parser.add_argument('--log_filename', type=str, help='log filename', default=None)
    args = parser.parse_args()


    env_id = args.env_id
    instance_number = args.instance_number
    load_models = args.load_models
    load_epoch = args.load_epoch
    scoring_method = args.scoring_method

    log_path = f'./logs/{scoring_method}/{env_id}/{str(instance_number)}'
    log_filename = os.path.join(log_path, 'log.txt')


    if not os.path.exists(log_path):
        os.makedirs(log_path)

    logging.basicConfig(level=logging.INFO, filename=log_filename, format='%(message)s')

    ## Creating Directory
    cwd = os.getcwd()
    directory = f'./checkpoints/{scoring_method}/{env_id}/{str(instance_number)}'
    directory = os.path.join(cwd, directory)
    if not os.path.exists(directory):
        os.makedirs(directory)

    newpath = os.path.join(cwd, env_id)
    newpath1 = os.path.join(newpath, str(instance_number))
    if not os.path.exists(newpath):
        os.makedirs(newpath)
    if not os.path.exists(newpath1):
        os.makedirs(newpath1)
    ## Making the agent with initial parameters
    env = gym.make(env_id)
    agent = Agent(alpha=3e-4, beta=3e-4, tau=0.002,
                  input_dims=env.observation_space.shape,
                  env=env, env_id=env_id, batch_size=256, layer1_size=256, layer2_size=256,
                  n_actions=env.action_space.shape[0], instance_number=instance_number,
                  checkpt_dir=directory, scoring_method=scoring_method)
    n_games = 50_000

    best_episode_reward = env.reward_range[0]
    step_reward_list = []
    episode_reward_list = []
    step_number = 0
    ev = 0
    save_every = 50
    reward_eval = []
    steps = []

    start_epoch = 0
    if load_models:
        logging.info(f'Loading trained models from {directory}')
        agent.load_models(load_epoch)
        start_epoch = load_epoch + 1
        step_reward_list = np.load(directory + '/s_r_' + env_id + '_' + str(instance_number) + '.npy').tolist()
        episode_reward_list = np.load(directory + '/ep_r_' + env_id + '_' + str(instance_number) + '.npy').tolist()
        steps = np.load(directory + '/t_s_' + env_id + '_' + str(instance_number) + '.npy', ).tolist()
        reward_eval = np.load(directory + '/eval_' + env_id + '_' + str(instance_number) + '.npy', ).tolist()

    for i in range(start_epoch, n_games):
        observation = env.reset()
        done = False
        ep_length = 0
        episode_reward = 0

        while not done:
            ep_length += 1
            # env.render(True)
            internal_state = env.env.sim.get_state()
            # To choose each action we have to use both observation and internal_state.
            # internal_state is then used to initialize the onboard model.
            action = agent.choose_action(observation, internal_state)
            observation_, reward, done, info = env.step(action)
            agent.remember(observation, action, reward, observation_, done)
            agent.learn()
            episode_reward += reward
            observation = observation_
            step_reward_list.append(reward)
            step_number += 1
            ## Eavaluating the agent per 1000 time-steps
            if step_number % 1000 == 0:
                env1 = gym.make(env_id)
                ep_r1 = 0
                observation1 = env1.reset()
                done1 = False
                ev += 1
                while not done1:
                    # observation1 = tf.convert_to_tensor([observation1], dtype=tf.float32)
                    internal_state1 = env1.env.sim.get_state()
                    # action1 = agent.actor.predict(observation1)
                    action1 = agent.choose_action(observation1, internal_state1, Test=True)
                    observation1, reward1, done1, _ = env1.step(action1)
                    ep_r1 += reward1
                reward_eval.append(ep_r1)
                logging.info(f'Eval {ev} : {ep_r1}')
        steps.append(step_number)
        episode_reward_list.append(episode_reward)
        average_episode_reward = np.mean(episode_reward_list[-100:])
        ## Saving agent if its performance gets better
        # if average_episode_reward > best_episode_reward:
        #     best_episode_reward = average_episode_reward
        #     agent.save_models(epoch=i, logger=logger)


        np.save(directory + '/last_eval_' + env_id + '_' + str(instance_number) + '.npy', reward_eval)

        if i % save_every == 0:
            logging.info(f'... saving models on epoch {i} ...')
            agent.save_models(epoch=i)
            np.save(directory + '/s_r_' + env_id + '_' + str(instance_number) + '.npy', step_reward_list)
            np.save(directory + '/ep_r_' + env_id + '_' + str(instance_number) + '.npy', episode_reward_list)
            np.save(directory + '/t_s_' + env_id + '_' + str(instance_number) + '.npy', steps)
            np.save(directory + '/eval_' + env_id + '_' + str(instance_number) + '.npy', reward_eval)
            gc.collect()
            keras.backend.clear_session()

        logging.info(f'episode {i} episode reward {episode_reward} average episode reward {average_episode_reward}')