-
Notifications
You must be signed in to change notification settings - Fork 0
/
sparse_utils.py
265 lines (187 loc) · 7.8 KB
/
sparse_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
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class Feather:
def __init__(self, gth, theta):
self.gth = gth
self.theta = theta
def forward(self, w):
return Feather_aux.apply(w, self.gth, self.theta)
class Feather_aux(torch.autograd.Function):
@staticmethod
def forward(ctx, w, gth, theta):
ctx.aux = torch.where(torch.abs(w) > gth, 1.0, theta)
p = 3
diff = torch.abs(w)**p - gth**p
w_masked = torch.where(diff > 0, torch.sign(w)*(diff)**(1/p), 0.0)
return w_masked
@staticmethod
def backward(ctx, g):
g = ctx.aux*g
return g, None, None
class SparseConv(nn.Module):
def __init__(self, conv, feather):
super(SparseConv, self).__init__()
self.conv = conv
self.feather = feather
def forward(self, x):
w = self.conv.weight
b = self.conv.bias
stride = self.conv.stride
padding = self.conv.padding
groups = self.conv.groups
if self.feather.gth > 0:
w = self.feather.forward(w)
out = F.conv2d(x, w, bias=b, padding=padding, stride=stride, groups=groups)
return out
class SparseFc(nn.Module):
def __init__(self, fc, feather):
super(SparseFc, self).__init__()
self.fc = fc
self.feather = feather
def forward(self, x):
w = self.fc.weight
b = self.fc.bias
if self.feather.gth > 0:
w = self.feather.forward(w)
out = F.linear(x, w, bias=b)
return out
def iter_sparsify(m, feather, pthres=0):
for name, child in m.named_children():
iter_sparsify(child, feather, pthres)
if type(child) == nn.Conv2d:
nw = (child.in_channels * child.out_channels * child.kernel_size[0] * child.kernel_size[1]) / child.groups
if nw >= pthres:
slayer = SparseConv(child, feather)
m.__setattr__(name, slayer)
if type(child) == nn.Linear:
nw = child.in_features * child.out_features
if nw >= pthres:
slayer = SparseFc(child, feather)
m.__setattr__(name, slayer)
def iter_desparsify(m, feather):
for name, child in m.named_children():
iter_desparsify(child, feather)
if type(child) == SparseConv:
conv = child.conv
w = conv.weight.data
nw = feather.forward(w)
conv.weight.data = nw
m.__setattr__(name, conv)
if type(child) == SparseFc:
fc= child.fc
w = fc.weight.data
nw = feather.forward(w)
fc.weight.data = nw
m.__setattr__(name, fc)
def get_params(model):
bn_ids =[]
modules = list(model.named_modules())
for n, layer in modules:
if isinstance(layer,torch.nn.modules.batchnorm.BatchNorm2d):
bn_ids.append(id(layer.weight))
bn_ids.append(id(layer.bias))
params, params_nowd = [], []
for name, p in model.named_parameters():
if id(p) in bn_ids or 'bias' in name:
params_nowd += [p]
else:
params += [p]
return params, params_nowd
def get_prunable_weights_cnt(model):
prunable_weights_cnt = 0
temp_dims = [0]
for name, layer in model.named_modules():
if ('Sparse' in layer.__class__.__name__):
if 'Conv' in layer.__class__.__name__ :
w = layer.conv.weight
elif 'Fc' in layer.__class__.__name__:
w = layer.fc.weight
else:
print(" Not Recognized Sparse Module ")
temp_dims.append(w.numel())
tnum = w.numel()
prunable_weights_cnt += tnum
idx_list = [0]
for i in range(len(temp_dims)):
idx_list.append(temp_dims[i] + idx_list[i])
return prunable_weights_cnt, idx_list
def calc_thresh(w, ratio):
w_sorted, _ = torch.sort(w)
m = (len(w_sorted)-1)*ratio
idx_floor, idx_ceil = int(np.floor(m)), int(np.ceil(m))
v1, v2 = w_sorted[idx_floor], w_sorted[idx_ceil]
thresh = v1 + (v2-v1)*(m-idx_floor)
return thresh.item()
def get_global_thresh(model, device, st_batch, prunable_weights_cnt, idx_list):
i = 1
w_total = torch.empty(prunable_weights_cnt).to(device)
for name, layer in model.named_modules():
if ('Sparse' in layer.__class__.__name__):
if 'Conv' in layer.__class__.__name__ :
w = layer.conv.weight.flatten().detach()
elif 'Fc' in layer.__class__.__name__:
w = layer.fc.weight.flatten().detach()
w_total[idx_list[i] : idx_list[i+1]] = w
i +=1
global_thresh = calc_thresh(torch.abs(w_total), st_batch)
return global_thresh
def pruning_scheduler(final_rate, nbatches, ntotalsteps, t1):
kf = final_rate
t1 = t1*nbatches
t2 = int(np.floor(ntotalsteps*0.5))
t = np.arange(t1,t2)
k = np.hstack(( np.zeros(t1), ( kf - kf*(1-(t-t1)/(t2-t1))**3), kf*np.ones(ntotalsteps-t2) ))
return k
def get_theta(final_rate):
if final_rate > 0.95:
theta = 0.5
else:
theta = 1.0
return theta
class Pruner:
def __init__(self, model, device, final_rate, nbatches, epochs, pthres=0, t1=0):
theta = get_theta(final_rate)
self.ntotalsteps = nbatches*epochs
self.step_idx = 0
self.feather = Feather(gth=0.0, theta=theta)
self.device = device
self.t1 = t1
self.model = model
iter_sparsify(m=self.model, feather=self.feather, pthres=pthres)
prunable_weights_cnt, idx_list = get_prunable_weights_cnt(self.model)
self.prunable_weights_cnt = prunable_weights_cnt
self.idx_list = idx_list
pscheduler = pruning_scheduler(final_rate, nbatches, self.ntotalsteps, self.t1)
self.pscheduler = pscheduler
def update_thresh(self, end_of_batch=False):
idx = self.step_idx
if end_of_batch:
idx -=1
st_batch = self.pscheduler[idx]
new_gth = 0.0
if st_batch > 0:
new_gth = get_global_thresh(self.model, self.device, st_batch, self.prunable_weights_cnt, self.idx_list)
self.feather.gth = new_gth
if not end_of_batch:
self.step_idx += 1
def print_sparsity(self):
local_zeros_cnt = 0
for name, layer in self.model.named_modules():
if ('Sparse' in layer.__class__.__name__):
if 'Conv' in layer.__class__.__name__ :
w = layer.conv.weight
elif 'Fc' in layer.__class__.__name__:
w = layer.fc.weight
else:
print(" Not Recognized Sparse Module ")
th = self.feather.gth
nw = F.hardshrink(w, th)
tsparsity = (nw == 0).float().sum().item()
local_zeros_cnt += tsparsity
tnum = nw.numel()
print(f'{name}'.ljust(40), f'#w: {int(tnum)}'.ljust(11), f'| sparsity: {round(100.0 * tsparsity / tnum, 2)}%'.ljust(18))
return 100 * float(local_zeros_cnt) / float(self.prunable_weights_cnt)
def desparsify(self):
iter_desparsify(self.model, self.feather)