/
train_instance.py
193 lines (172 loc) · 6.81 KB
/
train_instance.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
import time
import torch
torch.set_default_dtype(torch.float64)
import argparse
from inspect import signature
from lp import *
from layers import *
from gomory import *
from utilities import *
from train_utilities import *
SEED = 0
logger = configure_logging("train_log.txt")
class CutFunction(torch.nn.Module):
def __init__(self, input_size, nb_layers, nonlinear=False, size=32):
super().__init__()
self.inner_layers = SequentialSubadditive(
*[Cat(GomoryLayer(input_size + size*layer, size, nonlinear))
for layer in range(nb_layers)]
)
self.final_layer = LinearLayer(input_size+size*nb_layers, 1)
def forward(self, bias):
hidden = self.inner_layers(bias)
return self.final_layer(hidden).squeeze(0)
def train(instance_path, nb_layers=1, gomory_init=False, nonlinear=False, learning_rate=5e-4, target_noise=1e-4,
size=32, seed=0, gpu=0, add_variable_bounds=False, nb_steps=10000):
"""
Train a subadditive neural network to solve the subadditive dual of an instance.
Parameters
----------
instance_path: str or Path
Path to the instance.
nb_layers: int
Number of layers in the subadditive neural network.
gomory_init: bool
Should the neural network be initialized from the classical Gomory values?
nonlinear: bool
Should nonlinear Gomory cuts be used?
learning_rate: float
The learning rate of the optimization.
target_noise: float
How much noise should be added to the target in the algorithm?
size: int
Number of neurons (cuts) per layer.
seed: int
Seed to use.
gpu: int
Which gpu to use? (cpu=-1)
add_variable_bounds: bool
Should variable bounds be added to the problem?
nb_steps: int
For how many gradient steps to run the algorithm.
"""
short_path_name = Path(instance_path).parent.name + '/' + Path(instance_path).name
device = f'cuda:{gpu}' if gpu>=0 else 'cpu'
np.random.seed(seed), torch.manual_seed(seed)
A, b, c, vtypes, lp_value, lp_solution, ilp_value, \
ilp_solution, gomory_values = get_instance(instance_path, device=device, force_reload=True,
add_variable_bounds=add_variable_bounds)
cut_function = CutFunction(len(b), nb_layers, nonlinear, size).to(device)
if gomory_init:
gomory_initialization_(cut_function, A, b, c, vtypes)
optimizer = torch.optim.Adam(cut_function.parameters(), lr=learning_rate)
target, lower_bound = lp_solution, None
target_set = TensorSet()
time_start = time.perf_counter()
lower_bounds, is_step_lp, nb_targets, basis_start = [], [], [], None
for step in range(nb_steps):
extended_A, extended_b, c, vtypes = add_cuts_to_ilp(cut_function.inner_layers, A, b, c, vtypes)
gap = (extended_A@target - extended_b)[A.shape[0]:].min().item()
if step == 0 or gap < 1e-5:
# Normalize the cuts for the LP
cut_norms = extended_b[A.shape[0]:].abs() + 1e-5
cuts_A = extended_A[A.shape[0]:, :]/cut_norms.unsqueeze(-1)
cuts_b = extended_b[A.shape[0]:]/cut_norms
good_cuts = cuts_A.abs().max(-1).values > 1e-6
cuts_A, cuts_b = cuts_A[good_cuts, :], cuts_b[good_cuts]
lower_bound, target, basis_start = solve_lp(torch.cat([A, cuts_A]), torch.cat([b, cuts_b]), c,
basis_start=basis_start)
target_set.append(target)
is_step_lp.append(True)
else:
is_step_lp.append(False)
lower_bounds.append(lower_bound)
nb_targets.append(len(target_set))
noisy_target = (target + target_noise*torch.randn_like(target)).relu()
loss = extended_A@noisy_target - extended_b
optimizer.zero_grad()
loss.mean().backward()
optimizer.step()
if step % 100 == 0:
logger.info(f"[{'{i: >{n}}'.format(i=step, n=len(str(nb_steps)))}/{nb_steps}]"
f"[{short_path_name: <16}, {nb_layers}L"
f", {'G' if gomory_init else 'R'}, {'NL' if nonlinear else 'L': <2}]"
f" lp {lp_value: >6.2f}, nn {lower_bound: >6.2f}"
f", gomory {gomory_values[nb_layers]: >6.2f}, optimal {ilp_value: >6.2f}"
f" (gap {gap: >6.2f}, nb lps {np.sum(is_step_lp)})")
time_end = time.perf_counter()
solving_time = time_end - time_start
final_problem = (extended_A.detach().cpu(), extended_b.detach().cpu(), c.cpu(), vtypes)
return np.array(lower_bounds), np.array(is_step_lp), np.array(nb_targets), final_problem, ilp_value, solving_time
if __name__ == "__main__":
train_parameters = signature(train).parameters
parser = argparse.ArgumentParser()
parser.add_argument(
'instance_path',
help='Path to instance',
type=str,
)
parser.add_argument(
'-l', '--nb_layers',
help='Number of layers',
type=int,
default=train_parameters['nb_layers'].default,
)
parser.add_argument(
'-gi', '--gomory_init',
help='Should the layers be Gomory initialized?',
# type=bool,
action='store_true',
default=train_parameters['gomory_init'].default,
)
parser.add_argument(
'-nl', '--nonlinear',
help='Should nonlinear cuts be used?',
action='store_true',
default=train_parameters['nonlinear'].default,
)
parser.add_argument(
'-lr', '--learning_rate',
help='Learning rate',
type=float,
default=train_parameters['learning_rate'].default,
)
parser.add_argument(
'-tn', '--target_noise',
help='Target noise',
type=float,
default=train_parameters['target_noise'].default,
)
parser.add_argument(
'-sz', '--size',
help='Number of neurons (cuts) per layer',
type=int,
default=train_parameters['size'].default,
)
parser.add_argument(
'-s', '--seed',
help='Seed',
type=int,
default=train_parameters['seed'].default,
)
parser.add_argument(
'-g', '--gpu',
help='GPU (-1 for CPU)',
type=int,
default=train_parameters['gpu'].default,
)
parser.add_argument(
'-vb', '--add_variable_bounds',
help='Should variable bounds be added to the problem?',
# type=bool,
action='store_true',
default=train_parameters['add_variable_bounds'].default,
)
parser.add_argument(
'-ns', '--nb_steps',
help='Number of gradient steps to run the training',
type=int,
default=train_parameters['nb_steps'].default,
)
args = parser.parse_args()
train(**vars(args))