forked from AI4Finance-Foundation/ElegantRL
-
Notifications
You must be signed in to change notification settings - Fork 0
/
erl_run.py
154 lines (124 loc) · 6.35 KB
/
erl_run.py
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
142
143
144
145
146
147
148
149
150
151
152
153
154
import os
import time
from typing import List
import torch as th
import numpy as np
from erl_config import Config, build_env
from erl_agent import ReplayBuffer, AgentBase
def train_agent(args: Config):
args.init_before_training()
th.set_grad_enabled(False)
evaluator = Evaluator(
eval_env=build_env(args.env_class, args.env_args),
eval_per_step=args.eval_per_step,
eval_times=args.eval_times,
cwd=args.cwd,
)
env = build_env(args.env_class, args.env_args)
agent: AgentBase = args.agent_class(args.net_dims, args.state_dim, args.action_dim, gpu_id=args.gpu_id, args=args)
agent.last_state, info_dict = env.reset()
if args.if_off_policy:
buffer = ReplayBuffer(
gpu_id=args.gpu_id,
max_size=args.buffer_size,
state_dim=args.state_dim,
action_dim=1 if args.if_discrete else args.action_dim,
)
buffer_items = agent.explore_env(env, args.horizon_len * args.eval_times)
buffer.update(buffer_items) # warm up for ReplayBuffer
else:
buffer = []
'''start training'''
while True:
buffer_items = agent.explore_env(env, args.horizon_len)
if args.if_off_policy:
buffer.update(buffer_items)
else:
buffer[:] = buffer_items
logging_tuple = agent.update_net(buffer)
evaluator.evaluate_and_save(agent.act, args.horizon_len, logging_tuple)
if (evaluator.total_step > args.break_step) or os.path.exists(f"{args.cwd}/stop"):
break # stop training when reach `break_step` or `mkdir cwd/stop`
evaluator.close()
def valid_agent(env_class, env_args: dict, net_dims: List[int], agent_class, actor_path: str, render_times: int = 8):
env = build_env(env_class, env_args)
state_dim = env_args['state_dim']
action_dim = env_args['action_dim']
agent = agent_class(net_dims, state_dim, action_dim, gpu_id=-1)
actor = agent.act
print(f"| render and load actor from: {actor_path}")
actor.load_state_dict(th.load(actor_path, map_location=lambda storage, loc: storage))
for i in range(render_times):
cumulative_reward, episode_step = get_rewards_and_steps(env, actor, if_render=True)
print(f"|{i:4} cumulative_reward {cumulative_reward:9.3f} episode_step {episode_step:5.0f}")
class Evaluator:
def __init__(self, eval_env, eval_per_step: int = 1e4, eval_times: int = 8, cwd: str = '.'):
self.cwd = cwd
self.env_eval = eval_env
self.eval_step = 0
self.total_step = 0
self.start_time = time.time()
self.eval_times = eval_times # number of times that get episodic cumulative return
self.eval_per_step = eval_per_step # evaluate the agent per training steps
self.recorder = []
print("| Evaluator:"
"\n| `step`: Number of samples, or total training steps, or running times of `env.step()`."
"\n| `time`: Time spent from the start of training to this moment."
"\n| `avgR`: Average value of cumulative rewards, which is the sum of rewards in an episode."
"\n| `stdR`: Standard dev of cumulative rewards, which is the sum of rewards in an episode."
"\n| `avgS`: Average of steps in an episode."
"\n| `objC`: Objective of Critic network. Or call it loss function of critic network."
"\n| `objA`: Objective of Actor network. It is the average Q value of the critic network."
f"\n| {'step':>8} {'time':>8} | {'avgR':>8} {'stdR':>6} {'avgS':>6} | {'objC':>8} {'objA':>8}")
def evaluate_and_save(self, actor, horizon_len: int, logging_tuple: tuple):
self.total_step += horizon_len
if self.eval_step + self.eval_per_step > self.total_step:
return
self.eval_step = self.total_step
rewards_steps_ary = [get_rewards_and_steps(self.env_eval, actor) for _ in range(self.eval_times)]
rewards_steps_ary = np.array(rewards_steps_ary, dtype=np.float32)
avg_r = rewards_steps_ary[:, 0].mean() # average of cumulative rewards
std_r = rewards_steps_ary[:, 0].std() # std of cumulative rewards
avg_s = rewards_steps_ary[:, 1].mean() # average of steps in an episode
used_time = time.time() - self.start_time
self.recorder.append((self.total_step, used_time, avg_r))
save_path = f"{self.cwd}/actor_{self.total_step:012.0f}_{used_time:08.0f}_{avg_r:08.2f}.pth"
th.save(actor.state_dict(), save_path)
print(f"| {self.total_step:8.2e} {used_time:8.0f} "
f"| {avg_r:8.2f} {std_r:6.2f} {avg_s:6.0f} "
f"| {logging_tuple[0]:8.2f} {logging_tuple[1]:8.2f}")
def close(self):
np.save(f"{self.cwd}/recorder.npy", np.array(self.recorder))
draw_learning_curve_using_recorder(self.cwd)
def get_rewards_and_steps(env, actor, if_render: bool = False) -> (float, int):
device = next(actor.parameters()).device # net.parameters() is a Python generator.
state, info_dict = env.reset()
episode_steps = 0
cumulative_returns = 0.0 # sum of rewards in an episode
for episode_steps in range(12345):
tensor_state = th.as_tensor(state, dtype=th.float32, device=device).unsqueeze(0)
tensor_action = actor(tensor_state)
action = tensor_action.detach().cpu().numpy()[0] # not need detach(), because using torch.no_grad() outside
state, reward, terminated, truncated, _ = env.step(action)
cumulative_returns += reward
if if_render:
env.render()
if terminated or truncated:
break
cumulative_returns = getattr(env.unwrapped, 'cumulative_returns', cumulative_returns)
return cumulative_returns, episode_steps + 1
def draw_learning_curve_using_recorder(cwd: str):
recorder = np.load(f"{cwd}/recorder.npy")
import matplotlib as mpl
mpl.use('Agg') # write before `import matplotlib.pyplot as plt`. `plt.savefig()` without a running X server
import matplotlib.pyplot as plt
x_axis = recorder[:, 0]
y_axis = recorder[:, 2]
plt.plot(x_axis, y_axis)
plt.xlabel('#samples (Steps)')
plt.ylabel('#Rewards (Score)')
plt.grid()
file_path = f"{cwd}/LearningCurve.jpg"
# plt.show() # if use `mpl.use('Agg')` to draw figures without GUI, then plt can't plt.show()
plt.savefig(file_path)
print(f"| Save learning curve in {file_path}")