-
Notifications
You must be signed in to change notification settings - Fork 51
/
meta_evaluator.py
167 lines (136 loc) · 6.55 KB
/
meta_evaluator.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
import tensorflow as tf
import numpy as np
import time
from utils import logger
class Trainer():
def __init__(self,algo,
env,
sampler,
sample_processor,
policy,
n_itr,
batch_size=500,
start_itr=0,
num_inner_grad_steps=3):
self.algo = algo
self.env = env
self.sampler = sampler
self.sampler_processor = sample_processor
self.policy = policy
self.n_itr = n_itr
self.start_itr = start_itr
self.num_inner_grad_steps = num_inner_grad_steps
self.batch_size = batch_size
def train(self):
"""
Implement the repilte algorithm for ppo reinforcement learning
"""
start_time = time.time()
avg_ret = []
avg_pg_loss = []
avg_vf_loss = []
avg_latencies = []
for itr in range(self.start_itr, self.n_itr):
itr_start_time = time.time()
logger.log("\n ---------------- Iteration %d ----------------" % itr)
logger.log("Sampling set of tasks/goals for this meta-batch...")
paths = self.sampler.obtain_samples(log=True, log_prefix='')
""" ----------------- Processing Samples ---------------------"""
logger.log("Processing samples...")
samples_data = self.sampler_processor.process_samples(paths, log='all', log_prefix='')
""" ------------------- Inner Policy Update --------------------"""
policy_losses, value_losses = self.algo.UpdatePPOTarget(samples_data, batch_size=self.batch_size)
#print("task losses: ", losses)
print("average policy losses: ", np.mean(policy_losses))
avg_pg_loss.append(np.mean(policy_losses))
print("average value losses: ", np.mean(value_losses))
avg_vf_loss.append(np.mean(value_losses))
""" ------------------- Logging Stuff --------------------------"""
ret = np.sum(samples_data['rewards'], axis=-1)
avg_reward = np.mean(ret)
latency = samples_data['finish_time']
avg_latency = np.mean(latency)
avg_latencies.append(avg_latency)
logger.logkv('Itr', itr)
logger.logkv('Average reward, ', avg_reward)
logger.logkv('Average latency,', avg_latency)
logger.dumpkvs()
avg_ret.append(avg_reward)
return avg_ret, avg_pg_loss,avg_vf_loss, avg_latencies
if __name__ == "__main__":
from env.mec_offloaing_envs.offloading_env import Resources
from env.mec_offloaing_envs.offloading_env import OffloadingEnvironment
from policies.meta_seq2seq_policy import Seq2SeqPolicy
from samplers.seq2seq_sampler import Seq2SeqSampler
from samplers.seq2seq_sampler_process import Seq2SeSamplerProcessor
from baselines.vf_baseline import ValueFunctionBaseline
from meta_algos.ppo_offloading import PPO
from utils import utils, logger
logger.configure(dir="./meta_evaluate_ppo_log/task_offloading", format_strs=['stdout', 'log', 'csv'])
resource_cluster = Resources(mec_process_capable=(10.0 * 1024 * 1024),
mobile_process_capable=(1.0 * 1024 * 1024),
bandwidth_up=7.0, bandwidth_dl=7.0)
env = OffloadingEnvironment(resource_cluster=resource_cluster,
batch_size=100,
graph_number=100,
graph_file_paths=[
"./env/mec_offloaing_envs/data/meta_offloading_20/offload_random20_12/random.20."
],
time_major=False)
print("calculate baseline solution======")
env.set_task(0)
action, finish_time = env.greedy_solution()
target_batch, task_finish_time_batch = env.get_reward_batch_step_by_step(action[env.task_id],
env.task_graphs_batchs[env.task_id],
env.max_running_time_batchs[env.task_id],
env.min_running_time_batchs[env.task_id])
discounted_reward = []
for reward_path in target_batch:
discounted_reward.append(utils.discount_cumsum(reward_path, 1.0)[0])
print("avg greedy solution: ", np.mean(discounted_reward))
print("avg greedy solution: ", np.mean(task_finish_time_batch))
print("avg greedy solution: ", np.mean(finish_time))
print()
finish_time, energy_cost = env.get_all_mec_execute_time()
print("avg all remote solution: ", np.mean(finish_time))
print()
finish_time, energy_cost = env.get_all_locally_execute_time()
print("avg all local solution: ", np.mean(finish_time))
policy = Seq2SeqPolicy(obs_dim=17,
encoder_units=128,
decoder_units=128,
vocab_size=2,
name="core_policy")
sampler = Seq2SeqSampler(env,
policy,
rollouts_per_meta_task=1,
max_path_length=40000,
envs_per_task=None,
parallel=False)
baseline = ValueFunctionBaseline()
sample_processor = Seq2SeSamplerProcessor(baseline=baseline,
discount=0.99,
gae_lambda=0.95,
normalize_adv=True,
positive_adv=False)
algo = PPO(policy=policy,
meta_sampler=sampler,
meta_sampler_process=sample_processor,
lr=1e-4,
num_inner_grad_steps=3,
clip_value=0.2,
max_grad_norm=None)
# define the trainer of ppo to evaluate the performance of the trained meta policy for new tasks.
trainer = Trainer(algo=algo,
env=env,
sampler=sampler,
sample_processor=sample_processor,
policy=policy,
n_itr=21,
start_itr=0,
batch_size=500,
num_inner_grad_steps=3)
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
policy.load_variables(load_path="./meta_model_offload20_25_batch_10/meta_model_2900.ckpt")
avg_ret, avg_pg_loss, avg_vf_loss, avg_latencies = trainer.train()