-
Notifications
You must be signed in to change notification settings - Fork 4
/
spdy.py
261 lines (223 loc) · 8.51 KB
/
spdy.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
# SPDY search & DP algorithm implementation.
import copy
import math
import random
import numpy as np
import torch
from modelutils import *
class SPDY:
def __init__(
self,
target, db, errors, baselinetime, prunabletime, timings,
get_model, run, dataloader,
skip_layers=[], dpbuckets=10000
):
self.target = target
self.db = db
self.run = run
self.dpbuckets = dpbuckets
self.modelp = get_model()
self.layersp = find_layers(self.modelp)
self.batches = []
for batch in dataloader:
self.batches.append(run(self.modelp, batch, retmoved=True))
self.layers = list(db.layers())
self.layers = [l for l in self.layers if l not in skip_layers]
self.sparsities = [list(errors[self.layers[l]].keys()) for l in range(len(self.layers))]
self.costs = [
[errors[self.layers[l]][s] for s in self.sparsities[l]] for l in range(len(self.layers))
]
self.timings = [
[timings[self.layers[l]][s] for s in self.sparsities[l]] for l in range(len(self.layers))
]
self.baselinetime = baselinetime
self.prunabletime = prunabletime
targettime = self.baselinetime / self.target - (self.baselinetime - self.prunabletime)
best = sum(min(c) for c in self.timings)
if self.prunabletime < self.baselinetime:
print('Max target:', self.baselinetime / (best + self.baselinetime - self.prunabletime))
self.bucketsize = targettime / self.dpbuckets
for row in self.timings:
for i in range(len(row)):
row[i] = int(round(row[i] / self.bucketsize))
print('Loss/Base:', self.get_loss(self.modelp))
def dp(self, costs):
DP = np.full((len(costs), self.dpbuckets + 1), float('inf'))
PD = np.full((len(costs), self.dpbuckets + 1), -1)
for sparsity in range(len(costs[0])):
if costs[0][sparsity] < DP[0][self.timings[0][sparsity]]:
DP[0][self.timings[0][sparsity]] = costs[0][sparsity]
PD[0][self.timings[0][sparsity]] = sparsity
for layer in range(1, len(DP)):
for sparsity in range(len(costs[layer])):
timing = self.timings[layer][sparsity]
score = costs[layer][sparsity]
if timing == 0:
tmp = DP[layer - 1] + score
better = tmp < DP[layer]
if np.sum(better):
DP[layer][better] = tmp[better]
PD[layer][better] = sparsity
continue
if timing > self.dpbuckets:
continue
tmp = DP[layer - 1][:-timing] + score
better = tmp < DP[layer][timing:]
if np.sum(better):
DP[layer][timing:][better] = tmp[better]
PD[layer][timing:][better] = sparsity
score = np.min(DP[-1, :])
timing = np.argmin(DP[-1, :])
solution = []
for layer in range(len(DP) - 1, -1, -1):
solution.append(PD[layer][timing])
timing -= self.timings[layer][solution[-1]]
solution.reverse()
return solution
def gen_costs(self, coefs):
return [
[self.costs[i][j] * coefs[i] for j in range(len(self.costs[i]))] \
for i in range(len(self.costs))
]
def stitch_model(self, solution):
model = copy.deepcopy(self.modelp)
layers = find_layers(model)
config = {
self.layers[i]: self.sparsities[i][solution[i]] for i in range(len(self.layers))
}
self.db.stitch(layers, config)
return model
@torch.no_grad()
def get_loss(self, model):
loss = 0
for batch in self.batches:
loss += self.run(model, batch, loss=True)
return loss / len(self.batches)
def get_score(self, coefs):
costs = self.gen_costs(coefs)
solution = self.dp(costs)
model = self.stitch_model(solution)
return self.get_loss(model)
def save_profile(self, coefs, filename=''):
solution = self.dp(self.gen_costs(coefs))
if filename:
with open(filename, 'w') as f:
for i in range(len(solution)):
f.write('%s %s\n' % (self.sparsities[i][solution[i]], self.layers[i]))
else:
for i in range(len(solution)):
print('%s %s' % (self.sparsities[i][solution[i]], self.layers[i]))
def score(self, filename):
tmp = []
with open(filename, 'r') as f:
solution = []
for i, l in enumerate(f.readlines()):
splits = l.split(' ')
sparsity = splits[0]
tmp.append(float(sparsity))
name = splits[1][:-1]
j = self.sparsities[i].index(sparsity)
solution.append(j)
print('Speedup:', self.baselinetime / (
self.baselinetime - self.prunabletime + \
sum(t[s] for s, t in zip(solution, self.timings)) * self.bucketsize
))
model = self.stitch_model(solution)
print('Loss/Pruned:', self.get_loss(model))
return model
def dpsolve(self, save=''):
coefs = np.ones(len(self.layers))
print('Loss/Pruned:', self.get_score(coefs))
self.save_profile(coefs)
if save:
self.save_profile(coefs, save)
def search(
self, save='', randinits=100, maxnoimp=100, layerperc=.1
):
evals = 0
print('Finding init ...')
coefs = None
score = float('inf')
for _ in range(randinits):
coefs1 = np.random.uniform(0, 1, size=len(self.layers))
score1 = self.get_score(coefs1)
evals += 1
print('%04d %.4f %.4f' % (evals, score, score1))
if score1 < score:
score = score1
coefs = coefs1
print('Running local search ...')
for resamplings in range(round(layerperc * len(self.layers)), 0, -1):
print('Trying %d resamplings ...' % resamplings)
improved = True
while improved:
improved = False
for _ in range(maxnoimp):
coefs1 = coefs.copy()
for _ in range(resamplings):
coefs1[random.randint(0, len(self.layers) - 1)] = np.random.uniform(0, 1)
score1 = self.get_score(coefs1)
evals += 1
print('%04d %.4f %.4f' % (evals, score, score1))
if score1 < score:
score = score1
coefs = coefs1
improved = True
break
self.save_profile(coefs)
if save:
self.save_profile(coefs, save)
if __name__ == '__main__':
import argparse
from database import *
from datautils import *
parser = argparse.ArgumentParser()
parser.add_argument(
'model', type=str, choices=get_models,
help='Model to work with.'
)
parser.add_argument(
'dataset', type=str, choices=DEFAULT_PATHS,
help='Dataset to use.'
)
parser.add_argument(
'database', type=str,
help='Database location.'
)
parser.add_argument(
'timings', type=str,
help='Timings file.'
)
parser.add_argument(
'target', type=float,
help='Target speedup.'
)
parser.add_argument(
'profile', type=str,
help='Where to save the resulting profile.'
)
parser.add_argument(
'--datapath', type=str, default='',
help='Path to dataset.'
)
parser.add_argument(
'--seed', type=int, default=0,
help='Seed to use for calibration set selection.'
)
parser.add_argument(
'--nsamples', type=int, default=1024,
help='Number of samples in the calibration dataset.'
)
args = parser.parse_args()
get_model, test, run = get_functions(args.model)
dataloader, testloader = get_loaders(args.dataset, noaug=True, nsamples=args.nsamples)
model = get_model()
db = UnstrDatabase(args.database, model, skip=firstlast_names(args.model))
errors = db.get_errors()
baselinetime, prunabletime, timings = db.get_timings(args.timings)
spdy = SPDY(
args.target, db, errors, baselinetime, prunabletime, timings,
get_model, run, dataloader
)
PROFILE_FILE = args.profile
spdy.search(PROFILE_FILE)