-
Notifications
You must be signed in to change notification settings - Fork 21
Expand file tree
/
Copy pathREINFORCE_helper.py
More file actions
243 lines (213 loc) · 10.8 KB
/
REINFORCE_helper.py
File metadata and controls
243 lines (213 loc) · 10.8 KB
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
# Clase para cálculo de media y varianza de una secuencia
from time import time
import pandas as pd
import gym
import numpy as np
import moviepy.editor as mpy
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam, SGD
import keras.backend as K
from tensorboardX import SummaryWriter
import sklearn
import sklearn.preprocessing
def format_as_pandas(time_step, obs, preds, actions, rewards, disc_sum_rews, ep_returns, decimals = 3):
df = pd.DataFrame({'step': time_step.reshape(-1)})
df['observation'] = [np.array(r*10**decimals, dtype=int)/(10**decimals) for r in obs]
df['policy_distribution']=[np.array(r*10**decimals, dtype=int)/(10**decimals) for r in preds]
df['sampled_action'] = [np.array(r, dtype=int) for r in actions]
df['rewards']=rewards
df['discounted_sum_rewards']=np.array(disc_sum_rews*10**decimals, dtype=int)/(10**decimals)
df['episode_return']=np.array(ep_returns*10**decimals, dtype=int)/(10**decimals)
return df
class BaseAgent:
def __init__(self, ENV, logdir_root='logs', n_experience_episodes=1, gamma=0.999, epochs=1, lr=0.001, hidden_layer_neurons=128, EPISODES=2000, eval_period=50, algorithm='REINFORCE', noise=1.0, gif_to_board=False, fps=50, batch_size=128):
self.hidden_layer_neurons = hidden_layer_neurons
self.batch_size = batch_size
self.fps = fps
self.gif_to_board = gif_to_board
self.noise = noise
self.last_eval = 0
self.best_return = -np.inf
self.eval_period = 50
self.writer = None
self.epsilon = 1e-12
self.logdir_root = logdir_root
self.EPISODES = EPISODES
self.n_experience_episodes = n_experience_episodes
self.episode = 0
self.gamma = gamma
self.epochs = epochs
self.lr = lr
self.logdir = self.get_log_name(ENV, algorithm, logdir_root)
self.env = gym.make(ENV)
if type(self.env.action_space) != gym.spaces.box.Box:
self.nA = self.env.action_space.n
else:
print('Warning: El espacio de acción es continuo')
self.nA = self.env.action_space.shape[0]
self.logdir = self.logdir + '_' + str(self.noise)
if type(self.env.observation_space) == gym.spaces.box.Box:
self.nS = self.env.observation_space.shape[0]
else:
print('Warning: El espacio de observación no es continuo')
self.model = self.get_policy_model(lr=lr, hidden_layer_neurons=hidden_layer_neurons, input_shape=[self.nS] ,output_shape=self.nA)
state_space_samples = np.array(
[self.env.observation_space.sample() for x in range(10000)])
self.scaler = sklearn.preprocessing.StandardScaler()
self.scaler.fit(state_space_samples)
self.reset_env()
def get_policy_model(self, lr=0.001, hidden_layer_neurons = 128, input_shape=[4], output_shape=2):
pass
def get_log_name(self,ENV, algorithm, logdir_root):
name = logdir_root + '/'
name += ENV + '/' + algorithm + '/'
name += str(self.n_experience_episodes) + '_'
name += str(self.epochs) + '_'
name += str(self.batch_size) + '_'
name += str(self.gamma) + '_'
name += str(self.lr) + '_' + str(int(time()))
return name
def reset_env(self):
# Se suma uno a la cantidad de episodios
self.episode += 1
# Se observa el primer estado
self.observation = self.env.reset()
# Se resetea la lista con los rewards
self.reward = []
def get_experience_episodes(self, return_ts=False):
# Antes de llamar esta función hay que asegurarse de que el env esta reseteado
observations = []
actions = []
predictions = []
rewards = []
discounted_rewards = []
episodes_returns = []
episodes_lenghts = []
time_steps = []
exp_episodes = 0
ts_count = 0
# Juega n_experience_episodes episodios
while exp_episodes < self.n_experience_episodes:
# Obtengo acción
action, action_one_hot, prediction = self.get_action(eval=False)
# Ejecuto acción
observation, reward, done, info = self.env.step(action)
# Guardo reward obtenido por acción
self.reward.append(reward)
# Notar que se guarda la observación anterior
observations.append(self.observation)
actions.append(action_one_hot)
predictions.append(prediction.flatten())
rewards.append(reward)
self.observation = observation
ts_count+=1
time_steps.append(ts_count)
if done:
exp_episodes += 1
discounted_reward = self.get_discounted_rewards(self.reward)
discounted_rewards = np.hstack([discounted_rewards, discounted_reward])
ep_len = len(discounted_reward)
episodes_lenghts.append(ep_len)
episodes_returns = episodes_returns + [discounted_reward[0]]*ep_len
self.last_observation = self.observation
self.reset_env()
ts_count = 0
if return_ts:
return np.array(observations), np.array(actions), np.array(predictions), np.array(discounted_rewards), np.array(rewards), np.array(episodes_returns), np.array(episodes_lenghts), self.last_observation, np.array(time_steps).reshape(-1, 1)
else:
return np.array(observations), np.array(actions), np.array(predictions), np.array(discounted_rewards), np.array(rewards), np.array(episodes_returns), np.array(episodes_lenghts), self.last_observation
def log_data(self, episode, loss, ep_len_mean, entropy, rv, nomalized_loss, deltaT, ep_return, critic_loss=None, rv_normalized=None):
if self.writer is None:
self.writer = SummaryWriter(self.logdir)
print(f"correr en linea de comando: tensorboard --logdir {self.logdir_root}/")
print(f'\rEpisode: {episode}', end='')
self.writer.add_scalar('loss', loss, episode)
self.writer.add_scalar('episode_len', ep_len_mean, episode)
self.writer.add_scalar('entropy', entropy, episode)
self.writer.add_scalar('running_var', rv, episode)
self.writer.add_scalar('episode_return', ep_return, episode)
if rv_normalized:
self.writer.add_scalar('running_var_nomalized', rv_normalized, episode)
if nomalized_loss is not None:
self.writer.add_scalar('normalized_loss', nomalized_loss, episode)
self.writer.add_scalar('time', deltaT, episode)
if critic_loss is not None:
self.writer.add_scalar('critic_loss', critic_loss, episode)
if self.episode - self.last_eval >= self.eval_period:
if self.gif_to_board:
obs, actions, preds, disc_sum_rews, rewards, ep_returns, ep_len, frames = self.get_eval_episode(return_frames=self.gif_to_board)
else:
obs, actions, preds, disc_sum_rews, rewards, ep_returns, ep_len = self.get_eval_episode(return_frames=self.gif_to_board)
if self.best_return <= ep_returns[-1]:
self.model.save(self.logdir + '.hdf5')
print()
print(f'Model on episode {self.episode - 1} improved from {self.best_return} to {ep_returns[-1]}. Saved!')
self.best_return = ep_returns[-1]
if self.gif_to_board:
video = frames.reshape((1, )+frames.shape)
gif_name = str(self.episode) + '_' + self.logdir.replace('logs/', '').replace('/','_')
self.writer.add_video(gif_name, np.rollaxis(video, 4, 2), fps=self.fps)
self.writer.add_scalar('eval_episode_steps', len(obs), self.episode)
self.writer.add_scalar('eval_episode_return', ep_returns[-1], episode)
self.last_eval = self.episode
def get_eval_episode(self, gif_name=None, fps=50, return_frames=False):
frames=[]
self.reset_env()
observations = []
actions = []
predictions = []
rewards = []
discounted_rewards = []
episodes_returns = []
episodes_lenghts = []
exp_episodes = 0
if gif_name is not None or return_frames:
frames.append(self.env.render(mode = 'rgb_array'))
while True:
# Juega episodios hasta juntar un tamaño de buffer mínimo
action, action_one_hot, prediction = self.get_action(eval=True)
observation, reward, done, info = self.env.step(action)
self.reward.append(reward)
# Notar que se guarda la observación anterior
observations.append(self.observation)
actions.append(action_one_hot)
predictions.append(prediction.flatten())
rewards.append(reward)
self.observation = observation
if gif_name is not None or return_frames:
frames.append(self.env.render(mode = 'rgb_array'))
if done:
exp_episodes += 1
discounted_reward = self.get_discounted_rewards(self.reward)
discounted_rewards = np.hstack([discounted_rewards, discounted_reward])
ep_len = len(discounted_reward)
episodes_lenghts.append(ep_len)
episodes_returns = episodes_returns + [discounted_reward[0]]*ep_len
self.reset_env()
if gif_name is not None:
clip = mpy.ImageSequenceClip(frames, fps=fps)
clip.write_gif(gif_name, fps=fps, verbose=False, logger=None)
if return_frames:
return np.array(observations), np.array(actions), np.array(predictions), np.array(discounted_rewards), np.array(rewards), np.array(episodes_returns), np.array(episodes_lenghts), np.array(frames)
return np.array(observations), np.array(actions), np.array(predictions), np.array(discounted_rewards), np.array(rewards), np.array(episodes_returns), np.array(episodes_lenghts)
class RunningVariance:
# Keeps a running estimate of variance
def __init__(self):
self.m_k = None
self.s_k = None
self.k = None
def add(self, x):
if not self.m_k:
self.m_k = x
self.s_k = 0
self.k = 0
else:
old_mk = self.m_k
self.k += 1
self.m_k += (x - self.m_k) / self.k
self.s_k += (x - old_mk) * (x - self.m_k)
def get_variance(self, epsilon=1e-12):
return self.s_k / (self.k - 1 + epsilon) + epsilon
def get_mean(self):
return self.m_k