forked from eladhoffer/quantized.pytorch
-
Notifications
You must be signed in to change notification settings - Fork 1
/
reductions.py
290 lines (221 loc) · 10.1 KB
/
reductions.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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import torch as th
from utils.meters import AverageMeter
def spatial_mean(x):
return x.mean(tuple(range(2, x.dim()))) if x.dim() > 2 else x
def dim_reducer(x):
if x.dim() <= 3:
return x
return x.view(x.shape[0], x.shape[1], -1)
# k will ignore k-1 most extreme values
def spatial_edges(x,k=1,is_max=True,):
if x.dim() < 3:
return x
x_ = dim_reducer(x)
ret = x_.topk(k, -1, is_max)[0]
if is_max:
return ret[:,:,0]
return ret[:,:,k-1]
def spatial_min(x,k=1):
return spatial_edges(x,k,False)
def spatial_max(x,k=1):
return spatial_edges(x,k,True)
def spatial_margin(x,k=1):
return spatial_max(x,k)-spatial_min(x,k)
def spatial_min_max(x,k=1):
return th.stack([spatial_max(x,k),spatial_min(x,k)],1).view(x.shape[0],-1)
def spatial_mean_max(x,k=1):
return th.stack([spatial_max(x,k),spatial_mean(x)],1).view(x.shape[0],-1)
def spatial_min_mean_max(x, k=1):
return th.stack([spatial_max(x, k), spatial_mean(x), spatial_min(x, k)], 1).view(x.shape[0], -1)
def spatial_l2(x):
if x.dim() < 3:
return x
return th.norm(dim_reducer(x), dim=-1)
def max_local_response(x, grid=3):
if x.dim() < 3:
return x
if x.dim() < 4:
return th.nn.functional.adaptive_max_pool1d(x, grid).view(x.shape[0], x.shape[1], -1).log_softmax(-1).sum(
-1).neg_()
return th.nn.functional.adaptive_max_pool2d(x, grid).sum((-1, -2)).view(x.shape[0], x.shape[1], -1).log_softmax(
-1).sum(-1).neg_()
def sum_outer_product(x):
if x.dim() < 3:
return x
# r, c = th.tril_indices(x.shape[1], x.shape[1]).split(1)
x_ = x.view(x.shape[0], x.shape[1], -1)
x_ = x_.bmm(x_.transpose(2, 1)).log_softmax(-1).sum(-1).neg_()
# [:, :, r, c].view(x.shape[0],-1)
# x_=x_.sum(2)
return x_
def max_outer_product_flat(x):
if x.dim() < 3:
return x
r, c = th.tril_indices(x.shape[1], x.shape[1]).split(1)
x_ = x.view(x.shape[0], x.shape[1], -1).max(-1)[0].unsqueeze(2)
x_ = x_.bmm(x_.transpose(2, 1))[:, r, c].view(x.shape[0], -1)
#
# x_=x_.sum(2)
return x_
def G_p(ob, p=10):
temp = ob ** p
temp = temp.reshape(temp.shape[0], temp.shape[1], -1)
temp = ((th.matmul(temp, temp.transpose(dim0=2, dim1=1)))).sum(dim=2)
temp = (temp.sign() * th.abs(temp) ** (1 / p)).reshape(temp.shape[0], -1)
return temp
class MahalanobisDistance():
def __init__(self,mean,inv_cov):
self.mean = mean
self.inv_cov = inv_cov
self.use_mean_device = False
def __call__(self, x):
if x.device != self.mean.device:
# if not hasattr(self, 'use_mean_device') or self.use_mean_device != _USE_PERCENTILE_DEVICE:
# self.use_mean_device = _USE_PERCENTILE_DEVICE
if self.use_mean_device:
x = x.to(self.mean.device)
else:
self.mean = self.mean.to(x.device)
self.inv_cov = self.inv_cov.to(x.device)
x_c = x - self.mean
return (x_c.matmul(self.inv_cov).matmul(x_c.t())).diag().unsqueeze(1).sqrt()
class SumL1ChannelsDiff():
def __init__(self, mean):
self.mean = mean
self.use_mean_device = False
def __call__(self, x):
if x.device != self.mean.device:
# if not hasattr(self, 'use_mean_device') or self.use_mean_device != _USE_PERCENTILE_DEVICE:
# self.use_mean_device = _USE_PERCENTILE_DEVICE
if self.use_mean_device:
x = x.to(self.mean.device)
else:
self.mean = self.mean.to(x.device)
# return x.sub(self.mean).abs_().mean(1).unsqueeze(1)
return x.abs().mean(1).unsqueeze(1)
# clac Simes per batch element (samples x variables)
def calc_simes(pval):
pval, _ = th.sort(pval, 1)
view_shape = [-1, pval.shape[1]] + [1] * (pval.dim() - 2)
rank = th.arange(1, pval.shape[1] + 1, device=pval.device).view(view_shape)
simes_pval, _ = th.min(pval.shape[1] * pval / rank, 1)
return simes_pval.clamp(0,1).unsqueeze(1)
def calc_cond_fisher(pval, thresh=1):
pval[pval > thresh] = 1
return -2 * pval.log().sum(1).unsqueeze(1)
# rescaled fisher test
def calc_mean_fisher(pval):
return -2 * pval.log().mean(1).unsqueeze(1)
class StatefulReductionFactory():
def __init__(self, stateful_reduction, **reduction_kwargs):
self.reduction_ctor = stateful_reduction
self.reduction_kwargs = reduction_kwargs
self.reset()
def reset(self):
self.all_records = {}
self.curr_class = 0
self.curr_input = 0
self.set_class_input(0, 0)
def set_class(self, c, create_new=True):
self.curr_class = c
self.set_class_input(self.curr_class, self.curr_input, create_new)
def set_input(self, i, create_new=True):
self.curr_input = i
self.set_class_input(self.curr_class, self.curr_input, create_new)
def set_class_input(self, c, i, create_new=True):
self.class_input_specifier = (c, i)
if create_new and self.class_input_specifier not in self.all_records:
self.all_records[self.class_input_specifier] = {}
self.record = self.all_records[self.class_input_specifier]
def __call__(self, trace_name, create_new=True):
if create_new and trace_name not in self.record:
fn_ob = self.reduction_ctor(trace_name, **self.reduction_kwargs)
self.record[trace_name] = fn_ob
return self.record[trace_name]
class StatefulReduction():
def __init__(self, trace_name):
self.trace_name = trace_name
def __call__(self, x, measuring=False):
return self.specialized_reduction(self.shared_reduction(x), measuring)
def specialized_reduction(self, observations_list, measuring=False):
pass
def shared_reduction(self, x):
pass
class PdistanceGram(StatefulReduction):
def __init__(self, trace_name, precentile=0.01, powers=range(1, 11)):
super().__init__(trace_name)
self.precentile = precentile
self.powers = powers
self._quantile_high = [AverageMeter() for p in powers]
self._quantile_low = [AverageMeter() for p in powers]
def shared_reduction(self, x):
observations = []
for i in self.powers:
observations.append(G_p(x, i))
return observations
def specialized_reduction(self, observations_list, measuring=False):
l = 0
for e, i in enumerate(self.powers):
observations = observations_list[e]
if measuring:
tmp = observations.topk(int(self.precentile * observations.shape[0]), 0, True)[0]
p_95 = tmp[-1, :]
tmp = observations.topk(int(self.precentile * observations.shape[0]), 0, False)[0]
p_5 = tmp[-1, :]
self._quantile_high[e].update(p_95, observations.shape[0])
self._quantile_low[e].update(p_5, observations.shape[0])
if observations.device != self._quantile_high[e].mean.device:
self._quantile_high = self._quantile_high[e].mean.to(observations.device)
self._quantile_low = self._quantile_low[e].mean.to(observations.device)
l += (observations - self._quantile_high[e].mean).clamp(min=0) / self._quantile_high[e].mean.abs()
l += (self._quantile_low[e].mean - observations).clamp(min=0) / self._quantile_low[e].mean.abs()
return l
class Pdistance(StatefulReduction):
def __init__(self, trace_name, precentile=0.05, reduction_fn=spatial_max):
super().__init__(trace_name)
self.precentile = precentile
self.reduction_fn = reduction_fn
self._quantile_high = AverageMeter()
self._quantile_low = AverageMeter()
def shared_reduction(self, x):
observations = self.reduction_fn(x)
return observations
def specialized_reduction(self, observations, measuring=False):
if measuring:
tmp = observations.topk(int(self.precentile * observations.shape[0]), 0, True)[0]
p_95 = tmp[-1, :]
tmp = observations.topk(int(self.precentile * observations.shape[0]), 0, False)[0]
p_5 = tmp[-1, :]
self._quantile_high.update(p_95, observations.shape[0])
self._quantile_low.update(p_5, observations.shape[0])
if observations.device != self._quantile_high.mean.device:
self._quantile_high = self._quantile_high.mean.to(observations.device)
self._quantile_low = self._quantile_low.mean.to(observations.device)
l = (observations - self._quantile_high.mean).clamp(min=0) / self._quantile_high.mean.abs()
l += (self._quantile_low.mean - observations).clamp(min=0) / self._quantile_low.mean.abs()
return l
def __call__(self, x, measuring=False):
return self.specialized_reduction(self.shared_reduction(x), measuring)
# dev = (F.relu(mins[L][p] - g_p) / torch.abs(mins[L][p] + 1e-6)).sum(dim=1, keepdim=True)
# dev += (F.relu(g_p - maxs[L][p]) / torch.abs(maxs[L][p] + 1e-6)).sum(dim=1, keepdim=True)
_DEFAULT_SPATIAL_REDUCTIONS = {
'spatial-mean': spatial_mean,
# 'gram_p1':partial(G_p,p=1),
# 'gram_p2':partial(G_p,p=2),
# 'max_deviations_0.95': StatefulReductionFactory(Pdistance,**dict(precentile=0.05)),
# 'max_deviations_0.99': StatefulReductionFactory(Pdistance,**dict(precentile=0.01)),
# 'gram_div_0.95_p1':StatefulReductionFactory(PdistanceGram,**dict(powers=range(1,2),precentile=0.05)),
# 'gram_div_0.95_p2':StatefulReductionFactory(PdistanceGram,**dict(powers=range(1,3),precentile=0.05)),
# 'gram_div_0.95_p10':StatefulReductionFactory(PdistanceGram,**dict(powers=range(10,11),precentile=0.05)),
# 'gram_div_0.99_p1_p10':StatefulReductionFactory(PdistanceGram,**dict(precentile=0.01)),
'spatial-max': spatial_max,
# 'max_outer_product_flat':max_outer_product_flat,
# 'max_local_response' : max_local_response,
# 'sum_outer_product': sum_outer_product,
# 'spatial-min': spatial_min,
# 'spatial-min_max': spatial_min_max,
# 'spatial-mean_max': spatial_mean_max,
# 'spatial-min_mean_max': spatial_min_mean_max,
# 'spatial-l2':spatial_l2,
# 'spatial-l2':spatial_l2
}