-
Notifications
You must be signed in to change notification settings - Fork 0
/
energy_estimator.py
230 lines (194 loc) · 8.92 KB
/
energy_estimator.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
import argparse
import numpy as np
import torch
import torch.nn as nn
from erfnet_cp import erfnet
from torch.utils.data import DataLoader
import torch.nn.functional as F
import copy
from torch.nn.parameter import Parameter
import random
Alexnet_kernel_size = [11., 5., 3., 3., 3., 6., 1., 1.]
Alexnet_width_ub = [3, 64, 192, 384, 256, 256, 4096, 4096, 1000]
Mobilenet_kernel_size = [3., 3. / 32. + 1., 3. / 64 + 1., 3. / 128. + 1., 3. / 128. + 1., 3. / 256. + 1., 3. / 256. + 1.,
3. / 512. + 1., 3. / 512. + 1., 3. / 512. + 1., 3. / 512.+ 1., 3. / 512. + 1., 3. / 512. + 1.,
3. / 1024. + 1., 1.]
Mobilenet_width_ub = [3, 32, 64, 128, 128, 256, 256, 512, 512, 512, 512, 512, 512, 1024, 1024, 1001]
class EnergyEstimateWidthRescale(nn.Module):
def __init__(self, scales):
super(EnergyEstimateWidthRescale, self).__init__()
self.scales = Parameter(torch.tensor(scales, dtype=torch.float32), requires_grad=False)
def forward(self, x):
assert x.dim() != 1
x = x / self.scales
return torch.cat([(x[:, 0].detach() * x[:, 1]).unsqueeze(1),
x[:, 1:-2] * x[:, 2:-1],
(x[:, -2] * x[:, -1].detach()).unsqueeze(1)], dim=1)
class EnergyEstimateNet(nn.Module):
def __init__(self, n_nodes=None, preprocessor=None):
super(EnergyEstimateNet, self).__init__()
if n_nodes is None:
n_nodes = [len(Alexnet_width_ub) - 1, 1] # linear model for Alexnet
self.islinear = (len(n_nodes) == 2)
# self.preprocessor = EnergyEstimateWidthRescale([384.0] * 6 + [4096.0] * 3)
if preprocessor is not None:
self.preprocessor = preprocessor
else:
self.preprocessor = lambda x: x
layers = []
for i, _ in enumerate(n_nodes):
if i < len(n_nodes) - 1:
layer = nn.Linear(n_nodes[i], n_nodes[i + 1], bias=True)
if len(n_nodes) == 2:
layer.weight.data.zero_()
layer.bias.data.zero_()
layers.append(layer)
if i < len(n_nodes) - 2:
layers.append(nn.SELU())
self.regressor = nn.Sequential(*layers)
def forward(self, x):
single_data = (x.dim() == 1)
if single_data:
x = x.unsqueeze(0)
res = self.regressor(self.preprocessor(x))
if single_data:
res = res.squeeze(0)
return res
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Energy Estimator Training')
parser.add_argument('--infile', default='./energy_alexnet.npy', help='the input file of training data')
parser.add_argument('--outfile', default='./energymodel.pt', help='the output file of trained model')
parser.add_argument('--net', default='alexnet', help='network architecture')
parser.add_argument('--preprocess', default='rescale', help='preprocessor method')
parser.add_argument('--batch_size', type=int, default=-1, help='input batch size for training')
parser.add_argument('--seed', type=int, default=117, help='random seed (default: 117)')
parser.add_argument('--epochs', type=int, default=10000, help='number of epochs to train')
parser.add_argument('--wd', type=float, default=1e-3, help='weight decay')
parser.add_argument('--errhist', default=None, help='the output of error history')
parser.add_argument('--pinv', action='store_true', help='use pseudo inverse to solve (only for bilinear model)')
args = parser.parse_args()
print(args.__dict__)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# create data loader
data = np.load(args.infile)
np.random.shuffle(data)
val_portion = 0.2
val_num = round(data.shape[0] * val_portion)
tr_data, val_data = data[val_num:, :], data[:val_num, :]
preprocess = lambda x: x
tr_features, tr_labels = torch.from_numpy(preprocess(tr_data[:, :-2])), torch.from_numpy(tr_data[:, -2]).unsqueeze(
1)
val_features, val_labels = torch.from_numpy(preprocess(val_data[:, :-2])), torch.from_numpy(
val_data[:, -2]).unsqueeze(1)
if args.batch_size < 0:
args.batch_size = tr_features.size(0)
tr_loader = DataLoader(torch.utils.data.TensorDataset(tr_features, tr_labels), batch_size=args.batch_size,
shuffle=True)
val_loader = DataLoader(torch.utils.data.TensorDataset(val_features, val_labels), batch_size=args.batch_size,
shuffle=False)
def validate_model(model, loader, verbose=False):
with torch.no_grad():
model.eval()
mrae = 0.0
for data, label in loader:
data = data.cuda()
label = label.cuda()
output = model(data)
if verbose:
print(torch.cat([label, output], dim=-1))
# mae += F.l1_loss(label, output)
mrae += torch.mean(torch.abs(label - output) / torch.abs(label)).item()
mrae /= len(val_loader)
return mrae
if args.net == 'alexnet':
if args.preprocess == 'rescale':
model = EnergyEstimateNet(n_nodes=[len(Alexnet_width_ub) - 1, 1],
preprocessor=EnergyEstimateWidthRescale(scales=(Alexnet_width_ub)))
else:
raise NotImplementedError
elif args.net == 'mobilenet':
if args.preprocess == 'rescale':
model = EnergyEstimateNet(n_nodes=[len(Mobilenet_width_ub) - 1, 1],
preprocessor=EnergyEstimateWidthRescale(scales=(Mobilenet_width_ub)))
else:
raise NotImplementedError
elif args.net == 'erfnet':
if args.preprocess == 'rescale':
width_ub = [3] + erfnet().get_cpwub() + [20]
model = EnergyEstimateNet(n_nodes=[len(width_ub) - 1, 1],
preprocessor=EnergyEstimateWidthRescale(scales=(width_ub)))
else:
raise NotImplementedError
else:
raise NotImplementedError
if not args.pinv:
model.cuda()
# args.wd /= tr_features.shape[0]
optimizer = torch.optim.Adam(model.regressor.parameters(), lr=1e-3, weight_decay=args.wd)
# optimizer = torch.optim.SGD(model.regressor.parameters(), lr=1e-5, momentum=0.9, weight_decay=args.wd)
# optimizer = torch.optim.RMSprop(model.regressor.parameters(), lr=1e-2)
best_model = copy.deepcopy(model)
best_mrae = float('inf')
err_hist = []
err_res = []
if args.pinv:
assert model.islinear
model.cpu()
weight = None
bias = None
X = torch.cat([torch.ones((tr_features.shape[0], 1), dtype=tr_features.dtype),
model.preprocessor(tr_features).data], dim=1)
XtX = X.t().mm(X)
Y = tr_labels.data
XtY = X.t().mm(Y)
w = torch.gesv(XtY, XtX + 0.5 * args.wd * torch.eye(XtX.shape[0], dtype=XtX.dtype))[0].t()
print('linear system (XtX + 0.5(wd)I)w=XtY solved, w={}'.format(w))
for m in model.modules():
if isinstance(m, nn.Linear):
m.weight.data.copy_(w[:, 1:].data)
m.bias.data.copy_(w[:, 0].data)
break
model.cuda()
else:
for epoch in range(args.epochs):
model.train()
for data, label in tr_loader:
data = data.cuda()
label = label.cuda()
optimizer.zero_grad()
output = model(data)
mse = F.mse_loss(label, output)
# sys.stdout.write('{:.4e}===> '.format(mse))
mse.backward()
optimizer.step()
# sys.stdout.flush()
val_mrae = validate_model(model, loader=val_loader)
err_hist.append(val_mrae)
tr_mrae = 0#validate_model(model, loader=tr_loader)
if val_mrae < best_mrae:
best_model.load_state_dict(model.state_dict())
best_mrae = val_mrae
print("epoch {}, lr={}: tr_mrae={:.4e}, val_mrae={:.4e} (Best:{:.4e})"
.format(epoch, optimizer.param_groups[0]['lr'], tr_mrae, val_mrae, best_mrae))
print('val_mre={:.4e}'.format(validate_model(model, loader=val_loader)))
print(model.state_dict())
with torch.no_grad():
model.eval()
res = 0.0
for data, label in val_loader:
data = data.cuda()
label = label.cuda()
output = model(data)
for i in range(label.size(0)):
err_res.append((output[i].item(), label[i].item()))
torch.save(model.state_dict(), args.outfile)
if args.errhist is not None:
with open(args.errhist, 'w') as f:
for item in err_hist:
f.write("%s\n" % item)
with open(args.errhist + '.points', 'w') as f:
for item in err_res:
f.write("%s, %s\n" % (item[0], item[1]))