-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
242 lines (198 loc) · 9.89 KB
/
utils.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
import torch
from torch import nn
from torch.nn import functional as F
from matplotlib import pyplot as plt
from matplotlib_inline import backend_inline
from torch.optim import lr_scheduler
from dataload import load_cifar_10, load_cifar_100, load_tiny_imagenet
from network import ResNet20, Vgg, ResNet18
from pruner import Pruner
from train import train_net
def set_axes(axes, xlabel, ylabel, xlim, ylim, legend, xscale, yscale):
"""Set the axes for matplotlib.
Defined in :numref:`sec_calculus`"""
axes.set_xlabel(xlabel)
axes.set_ylabel(ylabel)
axes.set_xscale(xscale)
axes.set_yscale(yscale)
axes.set_xlim(xlim)
axes.set_ylim(ylim)
if legend:
axes.legend(legend)
axes.grid()
def plot(X, Y=None, xlabel=None, ylabel=None, legend=None, xlim=None, title=None,
ylim=None, xscale='linear', yscale='linear', xticklabels=None,
fmts=('^-r', '-b', '-g', '-m', '-C1', '-C5'), figsize=(4.5, 4), axes=None, twins=False, ylim2=None):
backend_inline.set_matplotlib_formats('svg')
plt.rcParams['figure.figsize'] = figsize
axes = axes if axes else plt.gca()
# Return True if `X` (tensor or list) has 1 axis
def has_one_axis(X):
return (hasattr(X, "ndim") and X.ndim == 1 or isinstance(X, list)
and not hasattr(X[0], "__len__"))
if has_one_axis(X):
X = [X]
if Y is None:
X, Y = [[]] * len(X), X
elif has_one_axis(Y):
Y = [Y]
if len(X) != len(Y):
X = X * len(Y)
axes.cla()
if twins:
ax2 = axes.twinx()
ax2.set_ylim(ylim2)
ax2.set_ylabel(ylabel[1])
i = 0
ax = axes
f = []
for x, y, fmt in zip(X, Y, fmts):
if twins and (i > 0):
ax = ax2
if len(x):
h, = ax.plot(x, y, fmt)
else:
h, = ax.plot(y, fmt)
f.append(h)
i += 1
if title:
axes.set_title(title)
set_axes(axes, xlabel, ylabel[0], xlim, ylim, legend, xscale, yscale)
if xticklabels:
axes.set_xticks(X[0])
axes.set_xticklabels(xticklabels, rotation=60, fontsize=12)
class Tester:
@staticmethod
def test_masked_resnet20(final_s, path="models/res", T=20, e=1.e-2):
epoch, batch_size, lr = 160, 128, 0.1
num_classes, prune_batch_size = 10, 256
loss = nn.CrossEntropyLoss().cuda(torch.device("cuda:0"))
data_iter = load_cifar_10(batch_size)
net = ResNet20().cuda(torch.device("cuda:0"))
resnet_pruner = Pruner(e, final_s)
resnet_pruner.prune(net, T, 5, (num_classes * 10, 3, 32, 32), prune_batch_size)
trainer = torch.optim.SGD(net.parameters(), weight_decay=1e-4, lr=lr, momentum=0.9)
scheduler = lr_scheduler.MultiStepLR(trainer, milestones=[80, 120], gamma=0.1)
sparsity_ratio, test_acc = train_net(net, loss, trainer, data_iter, epoch, path, scheduler)
print("true sparsity: %f%% test_acc: %.2f%%" % (sparsity_ratio * 100, test_acc * 100))
return sparsity_ratio, test_acc
@staticmethod
def test_masked_resnet20_muti(path="models/res"):
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
sparsity_ratios, test_accs = [], []
for final_s in sparsity:
sparsity_ratio, test_acc = Tester.test_masked_resnet20(final_s, path)
sparsity_ratios.append(sparsity_ratio)
test_accs.append(test_acc)
plot(compress_list, torch.tensor(test_accs) * 100, ylim=[76, 92],
xlabel='Sparsity (%)', ylabel=['Test Accuracy (%)'], title="ResNet-20 (CIFAR-10)",
xticklabels=[str(round(s * 100, 2)) for s in sparsity.tolist()], figsize=(6, 5))
plt.savefig("res/ResNet-20 (CIFAR-10)")
@staticmethod
def test_masked_vgg16(final_s, path="models/vgg16"):
epoch, batch_size, lr = 160, 128, 0.1
T, num_classes, prune_batch_size = 100, 100, 256
loss = nn.CrossEntropyLoss().cuda(torch.device("cuda:0"))
data_iter = load_cifar_100(batch_size)
net = Vgg(num_classes=num_classes).cuda(torch.device("cuda:0"))
resnet_pruner = Pruner(1.e-2, final_s)
resnet_pruner.prune(net, T, 5, (num_classes * 10, 3, 32, 32), prune_batch_size)
trainer = torch.optim.SGD(net.parameters(), weight_decay=5e-4, lr=lr, momentum=0.9, nesterov=True)
scheduler = lr_scheduler.MultiStepLR(trainer, milestones=[60, 120], gamma=0.1)
sparsity_ratio, test_acc = train_net(net, loss, trainer, data_iter, epoch, path, scheduler)
print("true sparsity: %f%% test_acc: %.2f%%" % (sparsity_ratio * 100, test_acc * 100))
return sparsity_ratio, test_acc
@staticmethod
def test_masked_vgg16_muti(path="models/vgg"):
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
sparsity_ratios, test_accs = [], []
for final_s in sparsity:
sparsity_ratio, test_acc = Tester.test_masked_vgg16(final_s, path)
sparsity_ratios.append(sparsity_ratio)
test_accs.append(test_acc)
plot(compress_list, torch.tensor(test_accs) * 100, #ylim=[66, 75],
xlabel='Sparsity (%)', ylabel=['Test Accuracy (%)'], title="VGG-16 (CIFAR-100)",
xticklabels=[str(round(s * 100, 2)) for s in sparsity.tolist()], figsize=(8, 8))
plt.savefig("res/VGG-16 (CIFAR-100)")
@staticmethod
def test_masked_resnet18(final_s, path="models/res18"):
epoch, batch_size, lr = 200, 256, 0.2
T, num_classes, prune_batch_size = 100, 200, 128 # 256太大
loss = nn.CrossEntropyLoss().cuda(torch.device("cuda:0"))
data_iter = load_tiny_imagenet(batch_size)
net = ResNet18().cuda(torch.device("cuda:0"))
resnet_pruner = Pruner(1.e-2, final_s)
resnet_pruner.prune(net, T, 5, (num_classes * 10, 3, 64, 64), prune_batch_size)
trainer = torch.optim.SGD(net.parameters(), weight_decay=1e-4, lr=lr, momentum=0.9)
scheduler = lr_scheduler.MultiStepLR(trainer, milestones=[100, 150], gamma=0.1)
sparsity_ratio, test_acc = train_net(net, loss, trainer, data_iter, epoch, path, scheduler)
print("true sparsity: %f%% test_acc: %.2f%%" % (sparsity_ratio * 100, test_acc * 100))
return sparsity_ratio, test_acc
@staticmethod
def test_masked_resnet18_muti(path="models/resnet18"):
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
sparsity_ratios, test_accs = [], []
for final_s in sparsity:
sparsity_ratio, test_acc = Tester.test_masked_resnet18(final_s, path)
sparsity_ratios.append(sparsity_ratio)
test_accs.append(test_acc)
plot(compress_list, torch.tensor(test_accs) * 100, # ylim=[66, 75],
xlabel='Sparsity (%)', ylabel=['Test Accuracy (%)'], title="res/ResNet-18 (Tiny-ImageNet)",
xticklabels=[str(round(s * 100, 2)) for s in sparsity.tolist()], figsize=(8, 8))
plt.savefig("res/ResNet-18(Tiny-ImageNet)")
@staticmethod
def compute_NTK():
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
T, num_classes, prune_batch_size = 20, 10, 256
for final_s in sparsity:
net = ResNet20().cuda(torch.device("cuda:0"))
resnet_pruner = Pruner(1.e-2, final_s)
dense_eigenvalues, sparse_eigenvalues = resnet_pruner.prune(net, T, 5, (num_classes * 10, 3, 32, 32),
prune_batch_size, NTK_show=True)
size = dense_eigenvalues.shape[0]
plot(torch.arange(size)+1, [dense_eigenvalues, sparse_eigenvalues],
legend=['Dense', 'NTK-SAP'], yscale='log', fmts=['-C7', '-r'],
xlabel='Eigenvalue index', ylabel=['Magnitude of each eigenvalue'], figsize=(8, 8))
plt.savefig(f"res/NTK_{int(final_s*100)}.png")
plot(torch.arange(size) + 1, [dense_eigenvalues.sort()[0], sparse_eigenvalues.sort()[0]],
legend=['Dense', 'NTK-SAP'], yscale='log', fmts=['-C7', '-r'],
xlabel='Eigenvalue index', ylabel=['Magnitude of each eigenvalue'], figsize=(8, 8))
plt.savefig(f"res/NTK_{int(final_s*100)}_sort.png")
@staticmethod
def test_masked_resnet20_muti_T(path="models/res"):
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
Ts = [1, 2, 4, 10, 20]
test_accs_T = []
for T in Ts:
test_accs = []
for final_s in sparsity:
_, test_acc = Tester.test_masked_resnet20(final_s, path, T=T)
test_accs.append(test_acc)
test_accs_T.append(test_accs)
plot(compress_list, torch.tensor(test_accs_T) * 100, xlabel='Sparsity (%)',
xticklabels=[str(round(s * 100, 2)) for s in sparsity.tolist()], figsize=(8, 8),
fmts=['o-m', 'o-y', 'o-g', 'o-c', 'o-b'],
ylabel=['Test Accuracy (%)'], legend=['T=1', 'T=2', 'T=4', 'T=10', 'T=20'])
plt.savefig("res/ResNet-20(CIFAR-10) muti_T")
@staticmethod
def test_masked_resnet20_muti_e(path="models/res"):
compress_list = torch.arange(2, 20, 2)
sparsity = 1 - 0.8 ** compress_list
epsilons = torch.sqrt(torch.tensor([1e-4, 2.5e-5, 1e-5, 1e-6]))
test_accs_T = []
for i, e in enumerate(epsilons):
test_accs = []
for final_s in sparsity:
_, test_acc = Tester.test_masked_resnet20(final_s, path, e=e)
test_accs.append(test_acc)
test_accs_T.append(test_accs)
plot(compress_list, torch.tensor(test_accs_T) * 100, xlabel='Sparsity (%)',
xticklabels=[str(round(s * 100, 2)) for s in sparsity.tolist()], figsize=(8, 8),
fmts=['o-m', 'o-y', 'o-g', 'o-c'],
ylabel=['Test Accuracy (%)'], legend=['ϵ=1e-4', 'ϵ=2.5e-5', 'ϵ=1e-5', 'ϵ=1e-6'])
plt.savefig("res/ResNet-20(CIFAR-10) muti_e")