forked from IntelLabs/networkgym
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
244 lines (217 loc) · 12.2 KB
/
train.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
#Copyright(C) 2023 Intel Corporation
#SPDX-License-Identifier: Apache-2.0
#File : start_client_sequential_training.py
# In this example, we perform sequential training for two environment sessions.
# the first session lasts for 3 episodes and the second session lasts for 1 episodes.
import numpy as np
import torch
import random
import sys
import os
import fire
import wandb
import copy
import time
from utils.buffer import ReplayBuffer
from CleanRL_agents.sac import SACAgent
from tqdm import tqdm
from network_gym_client import load_config_file
from network_gym_client import Env as NetworkGymEnv
from gymnasium.wrappers import NormalizeObservation
from utils.utils import *
sys.path.append('../')
sys.path.append('../../')
MODEL_SAVE_FREQ = 500
LOG_INTERVAL = 10
NUM_OF_EVALUATE_EPISODES = 10
EVAL_EPI_PER_SESSION = 1
EVAL_STEPS_PER_EPISODE = 50
#Each session will only have 10 episodes, and each episode will have 100 steps.
# When one of the client is terminated, the training program will sample a new client and continue training until the total number of steps reaches the num_steps
# For the candidate slice list, we can use the following slice list:
# 31 users is a boundary where the network is saturated, and we can observe dramatic delay difference between different slice list.
# we have different balanced slice list, and different unbalanced slice list.
slice_lists = slice_lists = [
[
{"num_users":6,"dedicated_rbg":0,"prioritized_rbg":12,"shared_rbg":25},
{"num_users":20,"dedicated_rbg":0,"prioritized_rbg":13,"shared_rbg":25},
{"num_users":5,"dedicated_rbg":0,"prioritized_rbg":0,"shared_rbg":25}
],
[
{"num_users":11,"dedicated_rbg":0,"prioritized_rbg":12,"shared_rbg":25},
{"num_users":15,"dedicated_rbg":0,"prioritized_rbg":13,"shared_rbg":25},
{"num_users":5,"dedicated_rbg":0,"prioritized_rbg":0,"shared_rbg":25}
],
[
{"num_users":13,"dedicated_rbg":0,"prioritized_rbg":12,"shared_rbg":25},
{"num_users":13,"dedicated_rbg":0,"prioritized_rbg":13,"shared_rbg":25},
{"num_users":5,"dedicated_rbg":0,"prioritized_rbg":0,"shared_rbg":25}
],
[
{"num_users":15,"dedicated_rbg":0,"prioritized_rbg":12,"shared_rbg":25},
{"num_users":11,"dedicated_rbg":0,"prioritized_rbg":13,"shared_rbg":25},
{"num_users":5,"dedicated_rbg":0,"prioritized_rbg":0,"shared_rbg":25}
],
[
{"num_users":20,"dedicated_rbg":0,"prioritized_rbg":12,"shared_rbg":25},
{"num_users":6,"dedicated_rbg":0,"prioritized_rbg":13,"shared_rbg":25},
{"num_users":5,"dedicated_rbg":0,"prioritized_rbg":0,"shared_rbg":25}
],
]
def main(agent_type:str,
env_name:str,
client_id = 0,
hidden_dim = 64,
steps_per_episode = 100,
episode_per_session = 5,
actor_lr = 1e-4,
critic_lr = 3e-4,
num_steps = 12000,
random_seed = 1
):
# client_id = 1
# env_name = "network_slicing"
storage_ver = 0
config_json = load_config_file(env_name)
init_list = random.sample(slice_lists, 1)
config_json["env_config"]["slice_list"] = init_list[0]
config_json["env_config"]["random_seed"] = random_seed
train_random_seed = random_seed
config_json["rl_config"]["agent"] = agent_type
config_json["env_config"]["steps_per_episode"] = steps_per_episode
config_json["env_config"]["episodes_per_session"] = episode_per_session
wandb.init(project = "network_gym_client", name = f"network_slicing_{agent_type}_training_{random_seed}", config = config_json)
# Create the environment
env = NetworkGymEnv(client_id, config_json, log=False) # make a network env using pass client id and configure file arguements.
# normalized_env = NormalizeObservation(env) # normalize the observation
normalized_env = env
num_steps = num_steps
# breakpoint()
obs, info = normalized_env.reset()
agent = SACAgent(state_dim=obs.shape[0],
action_dim=env.action_space.shape[0],
actor_lr=actor_lr,
critic_lr=critic_lr,
action_high=1,
action_low=0,
hidden_dim=hidden_dim)
if os.path.exists("./models/sac_model_{}_{}_{}_best.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed)):
best_model_path = "./models/sac_model_{}_{}_{}_best.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed)
agent.load(best_model_path) # load the best model, train from the best model.
eva_agent = copy.deepcopy(agent)
buffer = ReplayBuffer(max_size=1000000, obs_shape=obs.shape[0], n_actions=env.action_space.shape[0])
epsilon = 1.0
num_episodes = 0
progress_bar = tqdm(range(num_steps))
best_eval_reward = -np.inf
# Training loop
# Evaluate every 500 steps, same as model saving frequency
for step in progress_bar:
# if random.random() < epsilon:
# action = normalized_env.action_space.sample()
# else:
# breakpoint()
action = agent.predict(obs)
# if np.sum(action) >= 1: # Illegal action
# action = np.exp(action)/np.sum(np.exp(action))
action = np.exp(action)/np.sum(np.exp(action)) # softmax
# action = env.action_space.sample() # agent policy that uses the observation and info
nxt_obs, reward, terminated, truncated, info = normalized_env.step(action=action)
buffer.store(obs, action, reward, nxt_obs, truncated)
obs = nxt_obs
# breakpoint()
log = info_to_log(info)
wandb.log({"rewards/training_reward": reward})
wandb.log(log)
dataset = info_to_dataset(info)
buffer.store_raw_measurements(dataset)
if buffer.mem_cntr > 256:
training_batch = buffer.sample(256)
# breakpoint()
log_dict = agent.learn(*training_batch)
wandb.log(log_dict)
# If the environment is end, exit
if terminated:
print("Environment terminated, sampling a new environment...")
if num_steps - step < steps_per_episode * episode_per_session:
break
else:
print("Step: {}, Saving model...".format(step))
agent.save("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
eva_agent.load("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
eva_agent.actor.eval()
config_json["env_config"]["steps_per_episode"] = EVAL_STEPS_PER_EPISODE
config_json["env_config"]["episodes_per_session"] = EVAL_EPI_PER_SESSION
random_seed = 1
avg_reward = 0
for slice_list in slice_lists:
config_json["env_config"]["random_seed"] = random_seed
config_json["env_config"]["slice_list"] = slice_list
eval_env = NetworkGymEnv(client_id, config_json, log=False)
# normalized_eval_env = NormalizeObservation(eval_env)
normalized_eval_env = eval_env
env_reward, eval_log_dict = evaluate(eva_agent, normalized_eval_env, n_episodes=1)
avg_reward += env_reward
avg_reward /= len(slice_lists)
art = wandb.Artifact(f"{agent_type}-nn-{wandb.run.id}", type="model")
art.add_file("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
if avg_reward > best_eval_reward:
best_eval_reward = avg_reward
agent.save("./models/sac_model_{}_{}_{}_best.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed))
wandb.log_artifact(art, aliases=["latest", "best"])
else:
agent.save("./models/sac_model_{}_{}_{}_latest.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed))
wandb.log_artifact(art)
print("Step: {}, Eval Reward: {}".format(step, avg_reward))
wandb.log({"rewards/eval_reward": avg_reward})
buffer.save_buffer("./dataset/offline_train_buffer.h5")
buffer.save_raw_data("./dataset/offline_train_raw_data.h5")
storage_ver += 1
slice_list = random.sample(slice_lists, 1)
config_json["env_config"]["slice_list"] = slice_list[0]
config_json["env_config"]["random_seed"] = random.randint(0, 1000)
config_json["env_config"]["episodes_per_session"] = episode_per_session
env = NetworkGymEnv(client_id, config_json, log=False) # make a network env using pass client id and configure file arguements.
# normalized_env = NormalizeObservation(env) # normalize the observation
normalized_env = env
obs, info = normalized_env.reset()
print(f"New environment created, slice list: {slice_list[0]}, random seed: {config_json['env_config']['random_seed']}")
continue
# If the epsiode is up (environment still running), then start another one
if truncated:
obs, info = normalized_env.reset()
obs = torch.Tensor(obs)
epsilon = max(epsilon*0.95, 0.01)
num_episodes += 1
# if (step + 1) % MODEL_SAVE_FREQ == 0:
# print("Step: {}, Saving model...".format(step))
# agent.save("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
# eva_agent.load("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
# eva_agent.actor.eval()
# config_json["env_config"]["episodes_per_session"] = EVAL_EPI_PER_SESSION
# random_seed = 1
# avg_reward = 0
# for slice_list in slice_lists:
# config_json["env_config"]["random_seed"] = random_seed
# config_json["env_config"]["slice_list"] = slice_list
# eval_env = NetworkGymEnv(1, config_json, log=False)
# normalized_eval_env = NormalizeObservation(eval_env)
# env_reward = evaluate(eva_agent, normalized_eval_env, n_episodes=1)
# avg_reward += env_reward
# avg_reward /= len(slice_lists)
# art = wandb.Artifact(f"{agent_type}-nn-{wandb.run.id}", type="model")
# art.add_file("./models/sac_model_{}_{}_{}_ver{}.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed, storage_ver))
# if avg_reward > best_eval_reward:
# best_eval_reward = avg_reward
# agent.save("./models/sac_model_{}_{}_{}_best.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed))
# wandb.log_artifact(art, aliases=["latest", "best"])
# else:
# agent.save("./models/sac_model_{}_{}_{}_latest.ckpt".format(len(config_json["env_config"]["slice_list"]), config_json["rl_config"]["reward_type"], train_random_seed))
# wandb.log_artifact(art)
# print("Step: {}, Eval Reward: {}".format(step, avg_reward))
# wandb.log({"rewards/eval_reward": avg_reward})
# buffer.save_buffer("./dataset/offline_data_heavy_traffic_ver1.h5")
# storage_ver += 1
progress_bar.set_description("Step: {}, Reward: {:.3f}, Action: {}".format(step, reward, action))
if __name__ == "__main__":
fire.Fire(main)