-
Notifications
You must be signed in to change notification settings - Fork 25
/
Copy pathtrain.py
218 lines (181 loc) · 9.5 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
from torch.optim.lr_scheduler import LambdaLR
from mb_agg import *
from agent_utils import eval_actions
from agent_utils import select_action1,select_action2
from models.actor_critic import Job_Actor,Mch_Actor
from copy import deepcopy
import torch
import time
import torch.optim as optim
import torch.nn as nn
import numpy as np
from Params import configs
from validation import validate
from epsGreedyForMch import PredictMch
device = torch.device(configs.device)
from torch.utils.data import DataLoader
from uniform_instance import FJSPDataset
from FJSP_Env import FJSP
from rolloutbaseline import RolloutBaseline
import os
def adv_normalize(adv):
std = adv.std()
assert std != 0. and not torch.isnan(std), 'Need nonzero std'
n_advs = (adv - adv.mean()) / (adv.std() + 1e-8)
return n_advs
def train(epochs):
folder = 'FJSP-N_J-{}-N_M-{}'.format(configs.n_j, configs.n_m)
filename = 'rollout'
filepath = os.path.join(folder, filename)
g_pool_step = g_pool_cal(graph_pool_type=configs.graph_pool_type,
batch_size=torch.Size(
[configs.batch_size, configs.n_j * configs.n_m, configs.n_j * configs.n_m]),
n_nodes=configs.n_j * configs.n_m,
device=device)
job_actor = Job_Actor(n_j=configs.n_j,
n_m=configs.n_m,
num_layers=configs.num_layers,
learn_eps=False,
neighbor_pooling_type=configs.neighbor_pooling_type,
input_dim=configs.input_dim,
hidden_dim=configs.hidden_dim,
num_mlp_layers_feature_extract=configs.num_mlp_layers_feature_extract,
device=device)
mch_actor = Mch_Actor(n_j=configs.n_j,
n_m=configs.n_m,
num_layers=configs.num_layers,
learn_eps=False,
neighbor_pooling_type=configs.neighbor_pooling_type,
input_dim=configs.input_dim,
hidden_dim=configs.hidden_dim,
num_mlp_layers_feature_extract=configs.num_mlp_layers_feature_extract,
device=device).to(device)
validat_dataset = FJSPDataset(configs.n_j, configs.n_m, configs.low, configs.high,1280,200)
valid_loader = DataLoader(validat_dataset, batch_size=configs.batch_size)
rol_baseline = RolloutBaseline(job_actor,mch_actor, valid_loader, g_pool_step)
job_actor_optim = optim.Adam(job_actor.parameters(), lr=configs.lr)
mch_actor_optim = optim.Adam(mch_actor.parameters(), lr=configs.lr)
times, losses, rewards2, critic_rewards = [], [], [], []
start = time.time()
train_dataset = FJSPDataset(configs.n_j, configs.n_m, configs.low, configs.high, configs.num_ins,200)
plt = []
data_loader = DataLoader(train_dataset, batch_size=configs.batch_size)
for epoch in range(epochs):
job_actor.train()
mch_actor.train()
print("epoch:", epoch, "------------------------------------------------")
job_losses,mch_losses, rewards, critic_loss = [],[], [], []
for batch_idx, batch in enumerate(data_loader):
job_scheduler = LambdaLR(job_actor_optim, lr_lambda=lambda f: 0.98 ** batch_idx)
mch_scheduler = LambdaLR(mch_actor_optim, lr_lambda=lambda f: 0.98 ** batch_idx)
env = FJSP(configs.n_j, configs.n_m)
data = batch.numpy()
adj, fea, candidate, mask,mask_mch,dur,mch_time,job_time = env.reset(data)
job_log_prob = []
mch_log_prob = []
first_task = []
pretask = []
rewardssss = []
j = 0
job = candidate
hx = []
mch_node=[]
R_eward = []
actions = []
#env_mask_mch = torch.from_numpy(np.copy(mask_mch)).to(device)
while True:
#print(adj[0])
env_adj = aggr_obs(deepcopy(adj).to(device).to_sparse(), configs.n_j * configs.n_m)
env_mask_mch = torch.from_numpy(np.copy(mask_mch)).to(device)
env_fea = torch.from_numpy(np.copy(fea)).float().to(device)
env_fea = deepcopy(env_fea).reshape(-1, env_fea.size(-1))
env_candidate = torch.from_numpy(np.copy(candidate)).long().to(device)
env_dur = torch.from_numpy(np.copy(dur)).float().to(device)
env_mask = torch.from_numpy(np.copy(mask)).to(device)
env_mch_time = torch.from_numpy(np.copy(mch_time)).float().to(device)
env_job_time = torch.from_numpy(np.copy(job_time)).float().to(device)
action,log_a,action_node,_,mask_mch_action,hx= job_actor(x=env_fea,
graph_pool=g_pool_step,
padded_nei=None,
adj=env_adj,
candidate=env_candidate
,mask=env_mask
,pretask=pretask
,firsttask=first_task
,j=j
,mask_mch=env_mask_mch
,hx=hx
,mch_node=mch_node
,dur=env_dur
,job=job
,job_time=env_job_time
)
log_mch,mch_a,mch_node = mch_actor(action_node,hx,mask_mch_action,env_mch_time)
#print(action,mch_a)
job_log_prob.append(log_a.unsqueeze(1))
#print(action[0].item(),mch_a[0].item())
mch_log_prob.append(log_mch.unsqueeze(1))
if j == 0:
first_task = action.type(torch.long).to(device)
pretask =action.type(torch.long).to(device)
adj, fea, reward, done, candidate, mask,job,mask_mch,mch_time,job_time = env.step(action.cpu().numpy(), mch_a)
R_eward.append(reward)
if env.done():
break
job_log_p = torch.cat(job_log_prob, dim=1).sum(dim=1)
mch_log_p = torch.cat(mch_log_prob, dim=1).sum(dim=1)
reward = env.mchsEndTimes.max(-1).max(-1)
r_eward = torch.tensor(R_eward).permute(1,0).sum(-1)
base_reward = rol_baseline.eval(batch)
advantage = torch.tensor(reward - base_reward)
if (batch_idx + 1) % 1 == 0:
print('reward',reward.mean(),base_reward.mean())
print('advantage',torch.mean(advantage))
print('log_p',torch.mean(job_log_p).item())
print('log_p', torch.mean(mch_log_p).item())
advantage = adv_normalize(advantage)
job_actor_loss = torch.mean(advantage * job_log_p.cpu())
mch_actor_loss = torch.mean(advantage * mch_log_p.cpu())
job_actor_optim.zero_grad()
job_actor_loss.backward(retain_graph=True)
#grad_norms = clip_grad_norms(actor_optim.param_groups, 1)
#torch.nn.utils.clip_grad_norm_(actor.parameters(), 1)
job_actor_optim.step()
job_scheduler.step()
mch_actor_optim.zero_grad()
mch_actor_loss.backward(retain_graph=True)
# grad_norms = clip_grad_norms(actor_optim.param_groups, 1)
# torch.nn.utils.clip_grad_norm_(actor.parameters(), 1)
mch_actor_optim.step()
mch_scheduler.step()
rewards.append(np.mean(reward).item())
job_losses.append(torch.mean(job_actor_loss).item())
mch_losses.append(torch.mean(mch_actor_loss).item())
step = 10
if (batch_idx + 1) % step == 0:
end = time.time()
times.append(end - start)
start = end
job_mean_loss = np.mean(job_losses[-step:])
mch_mean_loss = np.mean(mch_losses[-step:])
mean_reward = np.mean(rewards[-step:])
print(' Batch %d/%d, reward: %2.3f, job_loss: %2.4f,mch_loss:%2.4f, took: %2.4fs' %
(batch_idx, len(data_loader), mean_reward, job_mean_loss,mch_mean_loss,
times[-1]))
if (batch_idx + 1) % 10 == 0:
rol_baseline.epoch_callback(job_actor, mch_actor, batch_idx)
print(plt)
rol_baseline.epoch_callback(job_actor,mch_actor, epoch)
epoch_dir = os.path.join(filepath, '%s' % epoch)
if not os.path.exists(epoch_dir):
os.makedirs(epoch_dir)
job_save_path = os.path.join(epoch_dir, 'job_actor.pt')
mch_save_path = os.path.join(epoch_dir, 'mch_actor.pt')
torch.save(job_actor.state_dict(), job_save_path)
torch.save(mch_actor.state_dict(), mch_save_path)
'''cost = rollout(actor, valid_loder, batch_size, n_nodes)
cost = cost.mean()
costs.append(cost.item())
print('Problem:TSP''%s' % n_nodes, '/ Average distance:', cost.item())
print(costs)'''
train(100)