-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_ppo_wandb.py
249 lines (200 loc) · 10.5 KB
/
train_ppo_wandb.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
import os
import sys
import glob
import time
import torch
import numpy as np
import pickle
from datetime import datetime
from parameters import configs
from environment.env import *
from policy import PPO, Memory
from instance_generator import one_instance_gen
from models.dag_aggregate import dag_pool
import wandb
# Example sweep configuration
count = 20
sweep_configuration = {
'method': 'bayes',
'name': 'sweep-v2',
'metric': {
'goal': 'maximize',
'name': 'reward'
},
'parameters': {
# 'batch_size': {'values': [16, 32, 64]},
'num_layers': {'values': [1, 2, 3, 4, 5]},
'num_mlp_layers_actor': {'values': [1, 2, 3, 4, 5]},
'num_mlp_layers_critic': {'values': [1, 2, 3, 4, 5]},
'k_epochs': {'values': [1, 2, 3]},
'eps_clip': {'max': 0.5, 'min': 0.01},
'lr_agent': {'max': 0.1, 'min': 0.0001},
'lr_critic': {'max': 0.1, 'min': 0.0001},
'entloss_coef': {'max': 0.5, 'min': 0.001},
'vloss_coef': {'values': [1, 2, 3]},
'ploss_coef': {'values': [1, 2, 3]}
}
}
sweep_id = wandb.sweep(sweep=sweep_configuration, project="project-DRL-AC")
device = torch.device(configs.device)
def main():
run = wandb.init()
print("Policy Case: ",configs.name)
print("\t + tasks: ",configs.n_tasks)
print("\t + devices: ",configs.n_devices)
print("\t + episodes: ",configs.max_updates)
configs.lr_agent = wandb.config.lr_agent
configs.lr_critic = wandb.config.lr_critic
configs.num_layers = wandb.config.num_layers
configs.num_mlp_layers_actor = wandb.config.num_mlp_layers_actor
configs.num_mlp_layers_critic = wandb.config.num_mlp_layers_critic
configs.k_epochs = wandb.config.k_epochs
configs.eps_clip = wandb.config.eps_clip
configs.lr_agent = wandb.config.lr_agent
configs.lr_critic = wandb.config.lr_critic
configs.entloss_coef = wandb.config.entloss_coef
configs.vloss_coef = wandb.config.vloss_coef
configs.ploss_coef = wandb.config.ploss_coef
#TODO clean old vars
torch.manual_seed(configs.torch_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(configs.torch_seed)
np.random.seed(configs.np_seed_train)
number_all_device_features = len(configs.feature_labels) #TODO fix
envs = [SPP(number_jobs=configs.n_jobs, number_devices=configs.n_devices,number_features=number_all_device_features) for _ in range(configs.num_envs)]
memories = [Memory() for _ in range(configs.num_envs)]
# initialize a PPO agent
ppo_agent = PPO(envs[0].state_dim)
# print(ppo_agent.policy)
dag_pool_step = dag_pool(graph_pool_type=configs.graph_pool_type,
batch_size=torch.Size([1, configs.n_tasks, configs.n_tasks]),
n_nodes=configs.n_tasks, device=device)
# training loop
log = []
logAlloc = []
for i_update in range(configs.max_updates):
#TODO clean vars -> state
ep_rewards = np.zeros(configs.num_envs)
init_rewards = np.zeros(configs.num_envs)
# alloc_envs = []
state_ft_envs,state_fm_envs= [],[]
candidate_envs = []
mask_envs = []
adj_envs = []
# Init all the environments
for i, env in enumerate(envs):
alloc, state, candidate, mask = env.reset(*one_instance_gen(n_jobs=configs.n_jobs, n_devices=configs.n_devices,cloud_features=configs.cloud_features, dependency_degree=configs.DAG_rand_dependencies_factor))
adj_envs.append(env.adj)
# alloc_envs.append(alloc)
state_ft_envs.append(state[0])
state_fm_envs.append(state[1])
candidate_envs.append(candidate)
mask_envs.append(mask)
ep_rewards[i] = - env.initQuality
init_rewards[i] = - env.initQuality
steps = 0
while True:
steps+=1
adj_tensor_envs = [torch.from_numpy(np.copy(adj)).to(device).to_sparse() for adj in adj_envs]
# alloc_tensor_envs = [torch.from_numpy(np.copy(alloc)).to(device) for alloc in alloc_envs]
state_ft_tensor_envs = [torch.from_numpy(np.copy(st)).to(device) for st in state_ft_envs]
state_fm_tensor_envs = [torch.from_numpy(np.copy(st)).to(device) for st in state_fm_envs]
candidate_tensor_envs = [torch.from_numpy(np.copy(candidate)).to(device) for candidate in candidate_envs]
mask_tensor_envs = [torch.from_numpy(np.copy(mask)).to(device) for mask in mask_envs]
with torch.no_grad():
task_action_envs,m_action_envs = [],[]
task_idx_envs, m_idx_envs = [],[]
for i in range(configs.num_envs):
# select action with policy
# state = torch.cat((feat_task_tensor_envs[i].reshape(-1),feat_mach_tensor_envs[i].reshape(-1)))
# state = state.type(torch.float)
task_action, ix_task_action, _, _, logProb, ix_machine_action, _, _, logProb_m = ppo_agent.policy_old(
state_ft=state_ft_tensor_envs[i],
state_fm=state_fm_tensor_envs[i].unsqueeze(0),
candidate=candidate_tensor_envs[i].unsqueeze(0),
mask=mask_tensor_envs[i].unsqueeze(0),
adj=adj_tensor_envs[i],
graph_pool=dag_pool_step)
# print(action)
# print(a_idx)
task_action_envs.append(task_action)
task_idx_envs.append(ix_task_action)
m_idx_envs.append(ix_machine_action)
memories[i].logprobs.append(logProb)
memories[i].logprobs_m.append(logProb_m)
# m_idx_envs.append(log_machprob)
# alloc_envs = []
state_ft_envs = []
state_fm_envs = []
# featT_envs = []
# featM_envs = []
candidate_envs = []
mask_envs = []
# Saving episode data
for i in range(configs.num_envs):
memories[i].adj_mb.append(adj_tensor_envs[i]) #TODO Purge memories
# memories[i].alloc_mb.append(alloc_tensor_envs[i])
memories[i].state_ft.append(state_ft_tensor_envs[i])
memories[i].state_fm.append(state_fm_tensor_envs[i])
# memories[i].featTask.append(feat_task_tensor_envs[i])
# memories[i].featMach.append(feat_mach_tensor_envs[i])
memories[i].candidate_mb.append(candidate_tensor_envs[i])
memories[i].mask_mb.append(mask_tensor_envs[i])
memories[i].a_mb.append(task_idx_envs[i]) #clean both vars.
memories[i].am_mb.append(m_idx_envs[i]) #clean both vars.
alloc, state, reward, done, candidate, mask = envs[i].step(task=int(task_action_envs[i]),
device=int(m_idx_envs[i]))
# alloc_envs.append(alloc)
state_ft_envs.append(state[0])
state_fm_envs.append(state[1])
# featT_envs.append(featTasks)
# featM_envs.append(featMachs)
candidate_envs.append(candidate)
mask_envs.append(mask)
ep_rewards[i] += reward
memories[i].reward_mb.append(reward)
memories[i].done_mb.append(done)
if envs[0].done(): #all environments are DONE (same number of tasks)
assert steps == envs[0].step_count
break
# if i_update in [0,5,10,20]:
if i_update in configs.record_alloc_episodes:
# print("Final placement: ",i_update)
# print(" -"*30)
for i in range(configs.num_envs): # Makespan
# print(i,envs[i].opIDsOnMchs,envs[i].feat_copy[envs[i].opIDsOnMchs][:,0],envs[i].feat_copy[envs[i].opIDsOnMchs][:,2])
logAlloc.append([i,envs[i].opIDsOnMchs.tolist(),envs[i].feat_copy[envs[i].opIDsOnMchs][:,0].tolist(),envs[i].feat_copy[envs[i].opIDsOnMchs][:,2].tolist()])
for j in range(configs.num_envs): # Makespan
ep_rewards[j] -= envs[j].posRewards # same actions/states as the initial maximum goal state
# update PPO agent
loss, v_loss = ppo_agent.update(memories)
for memory in memories:
memory.clear_memory()
mean_rewards_all_env = sum(ep_rewards) / len(ep_rewards)
mean_all_init_rewards = init_rewards.mean()
log.append([i_update, mean_rewards_all_env,v_loss,mean_all_init_rewards])
print('Episode {}\t Last reward: {:.2f}\t Mean_Vloss: {:.8f}\t Init reward: {:.2f}'.format(i_update + 1, mean_rewards_all_env, v_loss, mean_all_init_rewards))
wandb.log({"epoch":i_update + 1,"v_loss": v_loss, "loss": loss, "reward":mean_rewards_all_env, "init_reward":mean_all_init_rewards})
## DEBUG with out PPO Agent -.
# mean_rewards_all_env = ep_rewards.mean() # mean of the c-n time
# mean_all_init_rewards = init_rewards.mean()
# log.append([i_update, mean_rewards_all_env, mean_all_init_rewards])
# print('Episode {}\t Last reward: {:.2f} \t Init reward: {:.2f}'.format(i_update + 1, mean_rewards_all_env, mean_all_init_rewards))
#Store the logs
if configs.record_ppo:
with open('logs/log_ppo_' + str(configs.name) + "_" + str(configs.n_jobs) + '_' + str(configs.n_devices)+'.pkl', 'wb') as f:
pickle.dump(log, f)
if configs.record_alloc:
with open('logs/log_ppo_alloc_'+ str(configs.name) + "_" + str(configs.n_jobs) + '_' + str(configs.n_devices)+'.pkl', 'wb') as f:
pickle.dump(logAlloc, f)
wandb.finish()
print("Done\n")
if __name__ == '__main__':
print("TRAINING our policy in wandb plataform")
start_time = datetime.now().replace(microsecond=0)
print("Start training: ", start_time)
wandb.agent(sweep_id, function=main, count=count)
end_time = datetime.now().replace(microsecond=0)
print("Finish training: ", end_time)
print("Total time: ",(end_time-start_time))
print("Done policy test.")