-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_reinforce.py
152 lines (132 loc) · 6.68 KB
/
train_reinforce.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
import ecole
import torch
from pathlib import Path
from model import *
from utils import *
import numpy as np
import copy
import math
NB_EPOCHS = 10000
LEARNING_RATE = 0.0001
device_ids=range(torch.cuda.device_count())
torch.cuda.set_device('cuda:{}'.format(device_ids[0]))
# DEVICE = torch.device('cuda:1' if torch.cuda.is_available() else 'cpu')
check_path = "checkpoints/setcover/20210709_2344/para_best.pt"
scip_parameters = init_params(presolve=False)
# scip_parameters = {'separating/maxrounds': 0, 'presolving/maxrestarts': 0, 'limits/time': 3600}
env = ecole.environment.Branching(observation_function=ecole.observation.NodeBipartite(),
reward_function=ecole.reward.NNodes(),
information_function={"nodes": ecole.reward.NNodes().cumsum(),
"time": ecole.reward.SolvingTime().cumsum()},
scip_params=scip_parameters)
checkpoint = torch.load(check_path)
model = GNNPolicy()
model.cuda()
model.load_state_dict(checkpoint)
# model = torch.nn.DataParallel(model.cuda(), device_ids=device_ids)
# model.module.load_state_dict(checkpoint)
base_model = copy.deepcopy(model)
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
instances = ecole.instance.SetCoverGenerator(n_rows=500, n_cols=1000, density=0.05)
instances.seed(12)
env.seed(123)
Path("checkpoints/setcover/reinf").mkdir(parents=True, exist_ok=True)
def val_net(model, val_nums = 10):
instances_val = ecole.instance.SetCoverGenerator(n_rows=500, n_cols=1000, density=0.05)
instances_val.seed(111)
print(f"Val begin")
costs=[]
for instance_count, instance in zip(range(val_nums), instances_val):
observation, action_set, _, done, info_base = env.reset(instance)
while not done:
with torch.no_grad():
observation = (torch.from_numpy(observation.row_features.astype(np.float32)).cuda(),
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[0],
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[1],
torch.from_numpy(observation.edge_features.values.astype(np.float32)).view(-1, 1).cuda(),
torch.from_numpy(observation.column_features.astype(np.float32)).cuda())
logit = model(*observation)
logit = logit[action_set.astype(np.int64)]
prob = torch.softmax(logit, dim = -1)
observation, action_set, reward, done, info_base = env.step(action_set[prob.argmax()])
cost = info_base['nodes']
costs.append(cost)
return np.average(costs)
def train_net(log_probs, advantage):
log_probs = torch.stack(log_probs, 1)
# print(cost, baseline_v, ep)
loss = advantage * log_probs.mean()
# Perform backward pass and optimization step
optimizer.zero_grad()
loss.backward()
# Clip gradient norms and get (clipped) gradient norms for logging
grad_norms = clip_grad_norms(optimizer.param_groups, 10)
# print(advantage, grad_norms)
optimizer.step()
for instance_count, instance in zip(range(NB_EPOCHS), instances):
print(f"Instance: {instance_count} begin")
# run the baseline model
observation, action_set, _, done, info_base = env.reset(instance)
while not done:
with torch.no_grad():
observation = (torch.from_numpy(observation.row_features.astype(np.float32)).cuda(),
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[0],
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[1],
torch.from_numpy(observation.edge_features.values.astype(np.float32)).view(-1, 1).cuda(),
torch.from_numpy(observation.column_features.astype(np.float32)).cuda())
logit = base_model(*observation)
logit = logit[action_set.astype(np.int64)]
prob = torch.softmax(logit, dim = -1)
observation, action_set, reward, done, info_base = env.step(action_set[prob.argmax()])
baseline_v = info_base['nodes']
# Run the RL brancher
observation, action_set, _, done, info = env.reset(instance)
log_probs = []
max_per_update = 300
max_all_update = 3000
ep = 0
while not done:
ep += 1
with torch.set_grad_enabled(optimizer is not None):
observation = (torch.from_numpy(observation.row_features.astype(np.float32)).cuda(),
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[0],
torch.from_numpy(observation.edge_features.indices.astype(np.int64)).cuda()[1],
torch.from_numpy(observation.edge_features.values.astype(np.float32)).view(-1, 1).cuda(),
torch.from_numpy(observation.column_features.astype(np.float32)).cuda())
logit = model(*observation)
logit = logit[action_set.astype(np.int64)]
prob = torch.softmax(logit, dim = -1)
# Note that this is equivalent to what used to be called multinomial
m = torch.distributions.Categorical(prob)
action_ind = m.sample()
log_prob = m.log_prob(action_ind)
observation, action_set, reward, done, info = env.step(action_set[action_ind])
log_probs.append(log_prob.unsqueeze(0))
# print(ep, sep=",", end=",")
if ep > max_all_update:
baseline_v = info['nodes']
break
if ep > max_per_update:
# compute the gradient based on current cost and baseline
advantage = max(info['nodes'] - baseline_v, 0) / baseline_v
train_net(log_probs, advantage)
print(advantage, info['nodes'], baseline_v)
log_probs = []
# reset the baseline value to cost. The cost increase in the next iteration.
baseline_v = max(info['nodes'], baseline_v)
max_per_update += 300
if len(log_probs)==0:
continue
advantage = (info['nodes'] - baseline_v) / baseline_v
train_net(log_probs, advantage)
print(advantage, info['nodes'], baseline_v)
if instance_count % 100 == 0:
loss_current = val_net(model, 10)
loss_base = val_net(base_model, 10)
print("Val:", loss_current, loss_base)
if loss_current < loss_current:
base_model = model
print("!!!! Replace")
if instance_count % 500 == 0:
save_path = f"checkpoints/setcover/reinf/para_{instance_count}.pt"
torch.save(model.state_dict(), save_path)