-
Notifications
You must be signed in to change notification settings - Fork 11
/
Game.py
309 lines (261 loc) · 12.6 KB
/
Game.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
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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
'''
@File : Game.py
@Time : 2023/07/14 11:06:59
@Author : Zhou Zihao
@Version : 1.0
@Desc : None
'''
import os, json, sys
import gymnasium as gym
import numpy as np
import torch
import cv2
import time
import env
import algos
import skill
import utils
import cv2
from teacher_policy import TeacherPolicy
prefix = os.getcwd()
task_info_json = os.path.join(prefix, "prompt/task_info.json")
class Game:
def __init__(self, args, training=True):
# init seed
self.seed = args.seed
self.setup_seed(args.seed)
# init env
self.load_task_info(args.task, args.frame_stack, args.offline_planner, args.soft_planner)
# init logger
self.logger = utils.create_logger(args, training)
# init policy
if args.loaddir:
model_dir = os.path.join(args.logdir, args.policy, args.task, args.loaddir, args.loadmodel)
policy = torch.load(model_dir)
else:
policy = None
self.device = args.device
self.batch_size = args.batch_size
self.recurrent = args.recurrent
# self.student_policy = policy
self.student_policy = algos.PPO(policy,
self.obs_space,
self.action_space,
self.device,
self.logger.dir,
batch_size=self.batch_size,
recurrent=self.recurrent)
# init buffer
self.gamma = args.gamma
self.lam = args.lam
self.buffer = algos.Buffer(self.gamma, self.lam, self.device)
# other settings
self.n_itr = args.n_itr
self.traj_per_itr = args.traj_per_itr
self.num_eval = args.num_eval
self.eval_interval = args.eval_interval
self.save_interval = args.save_interval
self.total_steps = 0
def setup_seed(self, seed):
# setup seed for Numpy, Torch and LLM, not for env
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
torch.backends.cudnn.deterministic = True
def load_task_info(self, task, frame_stack, offline, soft):
print(f"[INFO]: resetting the task: {task}")
with open(task_info_json, 'r') as f:
task_info = json.load(f)
task = task.lower()
env_fn = utils.make_env_fn(task_info[task]['configurations'],
render_mode="rgb_array",
frame_stack = frame_stack)
self.env = utils.WrapEnv(env_fn)
self.obs_space = utils.get_obss_preprocessor(self.env.observation_space)[0]
self.action_space = self.env.action_space.n
self.max_ep_len = self.env.max_steps
prefix = task_info[task]['description'] + task_info[task]['example']
self.teacher_policy = TeacherPolicy(task, offline, soft, prefix, self.action_space, self.env.agent_view_size)
def train(self):
start_time = time.time()
for itr in range(self.n_itr):
print("********** Iteration {} ************".format(itr))
print("time elapsed: {:.2f} s".format(time.time() - start_time))
## collecting ##
sample_start = time.time()
self.buffer.clear()
n_traj = self.traj_per_itr
for _ in range(n_traj):
self.collect()
while len(self.buffer) < self.batch_size * 2:
self.collect()
n_traj += 1
total_steps = len(self.buffer)
samp_time = time.time() - sample_start
print("{:.2f} s to collect {:6n} timesteps | {:3.2f}sample/s.".format(samp_time, total_steps, (total_steps)/samp_time))
self.total_steps += total_steps
## training ##
optimizer_start = time.time()
mean_losses = self.student_policy.update_policy(self.buffer)
opt_time = time.time() - optimizer_start
try:
print("{:.2f} s to optimizer| loss {:6.3f}, entropy {:6.3f}, kickstarting {:6.3f}.".format(opt_time, mean_losses[0], mean_losses[1], mean_losses[2]))
except:
print(mean_losses)
## evaluate ##
if itr % self.eval_interval == 0 and itr > 0:
evaluate_start = time.time()
eval_returns = []
eval_lens = []
eval_success = []
for i in range(self.num_eval):
eval_outputs = self.evaluate(itr, record_frames=False)
eval_returns.append(eval_outputs[0])
eval_lens.append(eval_outputs[1])
eval_success.append(eval_outputs[2])
eval_time = time.time() - evaluate_start
print("{:.2f} s to evaluate.".format(eval_time))
if itr % self.save_interval == 0 and itr > 0:
self.student_policy.save(str(itr))
## log ##
if self.logger is not None:
avg_len = np.mean(self.buffer.ep_lens)
avg_reward = np.mean(self.buffer.ep_returns)
std_reward = np.std(self.buffer.ep_returns)
success_rate = sum(i > 0 for i in self.buffer.ep_returns) / n_traj
sys.stdout.write("-" * 49 + "\n")
sys.stdout.write("| %25s | %15s |" % ('Timesteps', self.total_steps) + "\n")
sys.stdout.write("| %25s | %15s |" % ('Return (train)', round(avg_reward,2)) + "\n")
sys.stdout.write("| %25s | %15s |" % ('Episode Length (train)', round(avg_len,2)) + "\n")
sys.stdout.write("| %25s | %15s |" % ('Success Rate (train)', round(success_rate,2)) + "\n")
if itr % self.eval_interval == 0 and itr > 0:
avg_eval_reward = np.mean(eval_returns)
avg_eval_len = np.mean(eval_lens)
eval_success_rate = np.sum(eval_success) / self.num_eval
sys.stdout.write("| %25s | %15s |" % ('Return (eval)', round(avg_eval_reward,2)) + "\n")
sys.stdout.write("| %25s | %15s |" % ('Episode Length (eval) ', round(avg_eval_len,2)) + "\n")
sys.stdout.write("| %25s | %15s |" % ('Success Rate (eval) ', round(eval_success_rate,2)) + "\n")
self.logger.add_scalar("Test/Return", avg_eval_reward, itr)
self.logger.add_scalar("Test/Eplen", avg_eval_len, itr)
self.logger.add_scalar("Test/Success Rate", eval_success_rate, itr)
sys.stdout.write("-" * 49 + "\n")
sys.stdout.flush()
self.logger.add_scalar("Train/Return Mean", avg_reward, itr)
self.logger.add_scalar("Train/Return Std", std_reward, itr)
self.logger.add_scalar("Train/Eplen", avg_len, itr)
self.logger.add_scalar("Train/Success Rate", success_rate, itr)
self.logger.add_scalar("Train/Loss", mean_losses[0], itr)
self.logger.add_scalar("Train/Mean Entropy", mean_losses[1], itr)
self.logger.add_scalar("Train/Kickstarting Loss", mean_losses[2], itr)
self.logger.add_scalar("Train/Policy Loss", mean_losses[3], itr)
self.logger.add_scalar("Train/Value Loss", mean_losses[4], itr)
self.logger.add_scalar("Train/Kickstarting Coef", self.student_policy.ks_coef, itr)
self.student_policy.save()
def collect(self):
'''
collect episodic data.
'''
with torch.no_grad():
obs = self.env.reset()
done = False
ep_len = 0
# reset student policy
mask = torch.FloatTensor([1]).to(self.device) # not done until episode ends
states = self.student_policy.model.init_states(self.device) if self.recurrent else None
# reset teacher policy
self.teacher_policy.reset()
while not done and ep_len < self.max_ep_len:
# get action from student policy
dist, value, states = self.student_policy(torch.Tensor(obs).to(self.device),
mask, states)
action = dist.sample()
log_probs = dist.log_prob(action)
action = action.to("cpu").numpy()
# get action from teacher policy
teacher_probs = self.teacher_policy(obs[0])
# interact with env
next_obs, reward, done, info = self.env.step(action)
# store in buffer
self.buffer.store(obs,
action,
reward,
value.to("cpu").numpy(),
log_probs.to("cpu").numpy(),
teacher_probs)
obs = next_obs
ep_len += 1
if done:
value = 0.
else:
value = self.student_policy(torch.Tensor(obs).to(self.device),
mask, states)[1].to("cpu").item()
self.buffer.finish_path(last_val=value)
def evaluate(self, itr=None, seed=None, record_frames=True, deterministic=False, teacher_policy=False):
with torch.no_grad():
# init env
seed = seed if seed else np.random.randint(1000000)
obs = self.env.reset(seed)
done = False
ep_len = 0
ep_return = 0.
if teacher_policy:
# init teacher policy
self.teacher_policy.reset()
else:
# init student policy
mask = torch.Tensor([1.]).to(self.device) # not done until episode ends
states = self.student_policy.model.init_states(self.device) if self.recurrent else None
# init vedio directory
if record_frames:
img_array = []
img = self.env.get_mask_render()
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img_array.append(img)
dir_name = 'teacher video' if teacher_policy else 'video'
dir_path = os.path.join(self.logger.dir, dir_name)
try:
os.makedirs(dir_path)
except OSError:
pass
while not done and ep_len < self.max_ep_len:
if teacher_policy:
# get action from teacher policy
probs = self.teacher_policy(obs[0])
if deterministic:
action = np.argmax(probs)
else:
action = np.random.choice(self.action_space, p=probs)
else:
# get action from student policy
dist, value, states = self.student_policy(torch.Tensor(obs).to(self.device), mask, states)
if deterministic:
action = torch.argmax(dist.probs).unsqueeze(0).to("cpu").numpy()
else:
action = dist.sample().to("cpu").numpy()
# interact with env
obs, reward, done, info = self.env.step(action)
ep_return += float(reward)
ep_len += 1
if record_frames:
img = self.env.get_mask_render()
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img_array.append(img)
ep_success = 1 if ep_return > 0 else 0
# save vedio
if record_frames:
height, width, layers = img.shape
size = (width,height)
video_name = "%s-%s.avi"%(itr, seed) if itr else "%s.avi"%seed
video_path = os.path.join(dir_path, video_name)
out = cv2.VideoWriter(video_path,
fourcc=cv2.VideoWriter_fourcc(*'DIVX'),
fps=3,
frameSize=size)
for img in img_array:
out.write(img)
out.release()
return ep_return, ep_len, ep_success
if __name__ == '__main__':
pass