/
driver.py
316 lines (276 loc) · 15.1 KB
/
driver.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
308
309
310
311
312
313
314
315
316
import copy
import torch
import torch.optim as optim
import torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
import ray
import os
import numpy as np
import random
import wandb
from attention import AttentionNet
from runner import RLRunner
from parameters import *
from env.task_env import TaskEnv
from scipy.stats import ttest_rel
ray.init()
writer = SummaryWriter(train_path)
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(gifs_path):
os.makedirs(gifs_path)
global_step = None
if WANDB_LOG:
wandb.init(project="CF")
def writeToTensorBoard(writer, tensorboardData, curr_episode, plotMeans=True):
# each row in tensorboardData represents an episode
# each column is a specific metric
if plotMeans == True:
tensorboardData = np.array(tensorboardData)
tensorboardData = list(np.nanmean(tensorboardData, axis=0))
reward, valueLoss, policyLoss, entropy, gradNorm, success_rate, time, time_cost, waiting, distance, effi = tensorboardData
else:
reward, valueLoss, policyLoss, entropy, gradNorm, success_rate, time, time_cost, waiting, distance, effi = tensorboardData
writer.add_scalar(tag='Losses/Policy Loss', scalar_value=policyLoss, global_step=curr_episode)
writer.add_scalar(tag='Losses/Entropy', scalar_value=entropy, global_step=curr_episode)
writer.add_scalar(tag='Losses/Grad Norm', scalar_value=gradNorm, global_step=curr_episode)
writer.add_scalar(tag='Losses/Value Loss', scalar_value=valueLoss, global_step=curr_episode)
writer.add_scalar(tag='Perf/Reward', scalar_value=reward, global_step=curr_episode)
writer.add_scalar(tag='Perf/Makespan', scalar_value=time, global_step=curr_episode)
writer.add_scalar(tag='Perf/Success rate', scalar_value=success_rate, global_step=curr_episode)
writer.add_scalar(tag='Perf/Time cost', scalar_value=time_cost, global_step=curr_episode)
writer.add_scalar(tag='Perf/Waiting time', scalar_value=waiting, global_step=curr_episode)
writer.add_scalar(tag='Perf/Traveling distance', scalar_value=distance, global_step=curr_episode)
writer.add_scalar(tag='Perf/Waiting Efficiency', scalar_value=effi, global_step=curr_episode)
if WANDB_LOG:
wandb.log({"Losses": {"Grad Norm": gradNorm, "Policy Loss": policyLoss, "Entropy": entropy},
"Perf": {"Reward": reward, "Time": time, "Success Rate": success_rate,
"Waiting Time": waiting, "Traveling Distance": distance, "Waiting Efficiency": effi}},
step=curr_episode)
def main():
device = torch.device('cuda') if USE_GPU_GLOBAL else torch.device('cpu')
local_device = torch.device('cuda') if USE_GPU else torch.device('cpu')
global_network = AttentionNet(AGENT_INPUT_DIM, TASK_INPUT_DIM, EMBEDDING_DIM).to(device)
baseline_network = AttentionNet(AGENT_INPUT_DIM, TASK_INPUT_DIM, EMBEDDING_DIM).to(device)
global_optimizer = optim.Adam(global_network.parameters(), lr=LR)
lr_decay = optim.lr_scheduler.StepLR(global_optimizer, step_size=DECAY_STEP, gamma=0.98)
# Automatically logs gradients of pytorch model
if WANDB_LOG:
wandb.watch(global_network)
curr_episode = 0
best_perf = -100
curr_level = 0
if LOAD_MODEL:
print('Loading Model...')
checkpoint = torch.load(model_path + '/checkpoint.pth')
global_network.load_state_dict(checkpoint['model'])
baseline_network.load_state_dict(checkpoint['model'])
global_optimizer.load_state_dict(checkpoint['optimizer'])
lr_decay.load_state_dict(checkpoint['lr_decay'])
curr_episode = checkpoint['episode']
curr_level = checkpoint['level']
print("curr_episode set to ", curr_episode)
if os.path.exists(model_path + '/best_model_checkpoint.pth'):
best_model_checkpoint = torch.load(model_path + '/best_model_checkpoint.pth')
best_perf = best_model_checkpoint['best_perf']
baseline_network.load_state_dict(best_model_checkpoint['model'])
print('best performance so far:', best_perf)
print(global_optimizer.state_dict()['param_groups'][0]['lr'])
if RESET_OPT:
global_optimizer = optim.Adam(global_network.parameters(), lr=LR)
lr_decay = optim.lr_scheduler.StepLR(global_optimizer, step_size=DECAY_STEP, gamma=0.98)
curr_episode = 0
# launch meta agents
meta_agents = [RLRunner.remote(i) for i in range(NUM_META_AGENT)]
# get initial weights
if device != local_device:
weights = global_network.to(local_device).state_dict()
baseline_weights = baseline_network.to(local_device).state_dict()
global_network.to(device)
baseline_network.to(device)
else:
weights = global_network.state_dict()
baseline_weights = baseline_network.state_dict()
# launch the first job on each runner
jobList = []
agents_num = np.random.randint(AGENTS_RANGE[0], AGENTS_RANGE[1] + 1)
tasks_num = np.random.randint(TASKS_RANGE[0], TASKS_RANGE[1] + 1)
for i, meta_agent in enumerate(meta_agents):
jobList.append(meta_agent.job.remote(weights, baseline_weights, curr_episode, agents_num, tasks_num))
curr_episode += 1
metric_name = ['success_rate', 'makespan', 'time_cost', 'waiting_time', 'travel_dist', 'efficiency']
tensorboardData = []
trainingData = []
experience_buffer = [[] for _ in range(9)]
test_set = np.random.randint(low=0, high=1e8, size=[256 // NUM_META_AGENT, NUM_META_AGENT])
baseline_value = None
try:
while True:
# wait for any job to be completed
done_id, jobList = ray.wait(jobList, num_returns=NUM_META_AGENT)
done_jobs = ray.get(done_id)
random.shuffle(done_jobs)
perf_metrics = {}
for n in metric_name:
perf_metrics[n] = []
for job in done_jobs:
jobResults, metrics, info = job
for i in range(9):
experience_buffer[i] += jobResults[i]
for n in metric_name:
perf_metrics[n].append(metrics[n])
update_done = False
while len(experience_buffer[0]) >= BATCH_SIZE:
agents_num = np.random.randint(AGENTS_RANGE[0], AGENTS_RANGE[1] + 1)
tasks_num = np.random.randint(TASKS_RANGE[0], TASKS_RANGE[1] + 1)
rollouts = copy.copy(experience_buffer)
for i in range(len(rollouts)):
rollouts[i] = rollouts[i][:BATCH_SIZE]
for i in range(len(experience_buffer)):
experience_buffer[i] = experience_buffer[i][BATCH_SIZE:]
if len(experience_buffer[0]) < BATCH_SIZE:
update_done = True
if update_done:
experience_buffer = []
for i in range(9):
experience_buffer.append([])
agent_inputs = torch.stack(rollouts[0], dim=0) # (batch,sample_size,2)
task_inputs = torch.stack(rollouts[1], dim=0) # (batch,sample_size,k_size)
action_batch = torch.stack(rollouts[2], dim=0) # (batch,1,1)
mask_batch = torch.stack(rollouts[3], dim=0) # (batch,1,1)
advantage_batch = torch.stack(rollouts[6], dim=0) # (batch,1,1)
reward_batch = torch.stack(rollouts[4], dim=0) # (batch,1,1)
index = torch.stack(rollouts[5])
if device != local_device:
agent_inputs = agent_inputs.to(device)
task_inputs = task_inputs.to(device)
action_batch = action_batch.to(device)
mask_batch = mask_batch.to(device)
reward_batch = reward_batch.to(device)
advantage_batch = advantage_batch.to(device)
index = index.to(device)
logp_list = global_network(task_inputs, agent_inputs, mask_batch) #, lstm_c, lstm_h)
logp = torch.gather(logp_list, 1, action_batch)
entropy = (logp_list * logp_list.exp()).nansum(dim=-1).mean()
policy_loss = - logp * advantage_batch.detach()
policy_loss = policy_loss.mean()
loss = policy_loss
global_optimizer.zero_grad()
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(global_network.parameters(), max_norm=10, norm_type=2)
global_optimizer.step()
lr_decay.step()
perf_data = []
for n in metric_name:
perf_data.append(np.nanmean(perf_metrics[n]))
data = [reward_batch.mean().item(), 0, policy_loss.item(),
entropy.item(), grad_norm.item(), *perf_data]
trainingData.append(data)
for i, meta_agent in enumerate(meta_agents):
jobList.append(meta_agent.job.remote(weights, baseline_weights, curr_episode, agents_num, tasks_num))
curr_episode += 1
if len(trainingData) >= SUMMARY_WINDOW:
writeToTensorBoard(writer, trainingData, curr_episode)
trainingData = []
# get the updated global weights
if update_done:
if device != local_device:
weights = global_network.to(local_device).state_dict()
baseline_weights = baseline_network.to(local_device).state_dict()
global_network.to(device)
baseline_network.to(device)
else:
weights = global_network.state_dict()
baseline_weights = baseline_network.state_dict()
if curr_episode % 512 == 0:
print('Saving model', end='\n')
checkpoint = {"model": global_network.state_dict(),
"optimizer": global_optimizer.state_dict(),
"episode": curr_episode,
"lr_decay": lr_decay.state_dict(),
"level": curr_level,
"best_perf": best_perf
}
path_checkpoint = "./" + model_path + "/checkpoint.pth"
torch.save(checkpoint, path_checkpoint)
print('Saved model', end='\n')
if EVALUATE:
if curr_episode % 1024 == 0:
# stop the training
ray.wait(jobList, num_returns=NUM_META_AGENT)
for a in meta_agents:
ray.kill(a)
torch.cuda.empty_cache()
print('Evaluate baseline model at ', curr_episode)
# test the baseline model on the new test set
if baseline_value is None:
test_agent_list = [RLRunner.remote(metaAgentID=i) for i in range(NUM_META_AGENT)]
for _, test_agent in enumerate(test_agent_list):
ray.get(test_agent.set_baseline_weights.remote(baseline_weights))
rewards = []
for i in range(256 // NUM_META_AGENT):
sample_job_list = []
for j, test_agent in enumerate(test_agent_list):
sample_job_list.append(test_agent.testing.remote(seed=test_set[i][j]))
sample_done_id, _ = ray.wait(sample_job_list, num_returns=NUM_META_AGENT)
reward = ray.get(sample_done_id)
rewards = rewards + reward
baseline_value = np.stack(rewards)
for a in test_agent_list:
ray.kill(a)
# test the current model's performance
test_agent_list = [RLRunner.remote(metaAgentID=i) for i in range(NUM_META_AGENT)]
for _, test_agent in enumerate(test_agent_list):
ray.get(test_agent.set_baseline_weights.remote(weights))
rewards = []
for i in range(256 // NUM_META_AGENT):
sample_job_list = []
for j, test_agent in enumerate(test_agent_list):
sample_job_list.append(test_agent.testing.remote(seed=test_set[i][j]))
sample_done_id, _ = ray.wait(sample_job_list, num_returns=NUM_META_AGENT)
reward = ray.get(sample_done_id)
rewards = rewards + reward
test_value = np.stack(rewards)
for a in test_agent_list:
ray.kill(a)
meta_agents = [RLRunner.remote(i) for i in range(NUM_META_AGENT)]
# update baseline if the model improved more than 5%
print('test value', test_value.mean())
print('baseline value', baseline_value.mean())
if test_value.mean() > baseline_value.mean():
_, p = ttest_rel(test_value, baseline_value)
print('p value', p)
if p < 0.05:
print('update baseline model at ', curr_episode)
if device != local_device:
weights = global_network.to(local_device).state_dict()
global_network.to(device)
else:
weights = global_network.state_dict()
baseline_weights = copy.deepcopy(weights)
baseline_network.load_state_dict(baseline_weights)
test_set = np.random.randint(low=0, high=1e8, size=[256 // NUM_META_AGENT, NUM_META_AGENT])
print('update test set')
baseline_value = None
best_perf = test_value.mean()
print('Saving best model', end='\n')
checkpoint = {"model": global_network.state_dict(),
"optimizer": global_optimizer.state_dict(),
"episode": curr_episode,
"lr_decay": lr_decay.state_dict(),
"best_perf": best_perf}
path_checkpoint = "./" + model_path + "/best_model_checkpoint.pth"
torch.save(checkpoint, path_checkpoint)
print('Saved model', end='\n')
jobList = []
for i, meta_agent in enumerate(meta_agents):
jobList.append(meta_agent.job.remote(weights, baseline_weights, curr_episode, agents_num, tasks_num))
curr_episode += 1
except KeyboardInterrupt:
print("CTRL_C pressed. Killing remote workers")
if WANDB_LOG:
wandb.finish()
for a in meta_agents:
ray.kill(a)
if __name__ == "__main__":
main()