-
Notifications
You must be signed in to change notification settings - Fork 4
/
adaprune.py
343 lines (297 loc) · 10.8 KB
/
adaprune.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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
# AdaPrune & global AdaPrune implementations for unstructured, blocked and N:M pruning.
import copy
import math
import torch
import torch.nn as nn
class MagLayerPruner:
# Assumes that all 0s have already been pruned
def __init__(self, layer, sparsity, lr=1e-3):
self.layer = layer
tmp = torch.sort(torch.abs(self.layer.weight.data.reshape(-1)))[0]
thresh = tmp[int(self.layer.weight.numel() * sparsity)]
self.mask = torch.abs(self.layer.weight.data) > thresh
self.apply_mask()
self.optim = torch.optim.Adam([self.layer.weight], lr=lr)
def optim_step(self, inp, out):
norm = torch.norm(out).item() ** 2
out1 = self.layer(inp)
out1.sub_(out)
out1.pow_(2)
loss = torch.sum(out1) / norm
loss.backward()
self.optim.step()
self.optim.zero_grad()
self.apply_mask()
def apply_mask(self):
self.layer.weight.data *= self.mask
class NM50LayerPruner(MagLayerPruner):
# Assume number of weights in layer is divisible by blocksize
def __init__(self, layer, blocksize, lr=1e-3):
self.layer = layer
w = self.layer.weight.data
if len(w.shape) == 4:
w = w.permute(0, 2, 3, 1)
_, i = torch.topk(
torch.abs(w.reshape((-1, blocksize))), blocksize // 2, dim=1
)
self.mask = torch.zeros_like(w).reshape(-1, blocksize)
for j in range(blocksize // 2):
self.mask[torch.arange(self.mask.shape[0]), i[:, j]] = 1
self.mask = self.mask.reshape(w.shape)
if len(w.shape) == 4:
self.mask = self.mask.permute(0, 3, 1, 2)
self.mask = self.mask == 1
self.apply_mask()
self.optim = torch.optim.Adam([self.layer.weight], lr=lr)
class BlockLayerPruner(MagLayerPruner):
# Assume number of weights in layer is divisible by blocksize
def __init__(self, layer, blocksize, sparsity, lr=1e-3):
self.layer = layer
w = self.layer.weight.data
if len(w.shape) == 4:
w = w.permute(0, 2, 3, 1)
tmp = torch.sum(torch.abs(w.reshape((-1, blocksize))), 1)
thresh = torch.sort(tmp)[0][int(tmp.numel() * sparsity)]
self.mask = torch.zeros_like(w).reshape(-1, blocksize)
self.mask[tmp > thresh, :] = 1
self.mask = self.mask.reshape(w.shape)
if len(w.shape) == 4:
self.mask = self.mask.permute(0, 3, 1, 2)
self.mask = self.mask == 1
self.apply_mask()
self.optim = torch.optim.Adam([self.layer.weight], lr=lr)
# Assume that we only prune the weight `parameter` of each layer
# Assume that `modelp` and `modeld` are on the same GPU
# Assume models are in eval mode
def layerw_adaprune(
pruners, modeld, dataloader, run, iters=10
):
layersd = find_layers(modeld)
def hook(name):
def tmp(layer, inp, out):
with torch.enable_grad():
pruners[name].optim_step(inp[0].data, out.data)
return tmp
handles = []
for name in pruners:
handles.append(layersd[name].register_forward_hook(hook(name)))
dev = layersd[next(iter(layersd))].weight.device
for i in range(iters):
print(i)
for batch in dataloader:
with torch.no_grad():
run(modeld, batch)
for h in handles:
h.remove()
def global_adaprune(
pruners, modelp, modeld, dataloader, run,
iters=100, lr=1e-5
):
layersp = find_layers(modelp)
layersd = find_layers(modeld)
def cache_output(name, outputs):
def tmp(layer, inp, out):
outputs[name] = out
return tmp
outputsp = {}
handlesp = []
for name in layersp:
handlesp.append(
layersp[name].register_forward_hook(cache_output(name, outputsp))
)
outputsd = {}
handlesd = []
for name in layersd:
handlesd.append(
layersd[name].register_forward_hook(cache_output(name, outputsd))
)
dev = layersp[next(iter(layersp))].weight.device
criterion = nn.MSELoss(reduction='sum')
optim = torch.optim.Adam(modelp.parameters(), lr=lr)
for i in range(iters):
cumloss = 0
for batch in dataloader:
with torch.no_grad():
run(modeld, batch)
run(modelp, batch)
loss = 0
for name in outputsd:
norm = torch.norm(outputsd[name].data).item() ** 2
loss += criterion(outputsp[name], outputsd[name].data) / norm
cumloss += loss.item()
loss.backward()
optim.step()
optim.zero_grad()
for p in pruners.values():
p.apply_mask()
print('%05d: %.6f' % (i, cumloss / len(dataloader)))
for h in handlesp:
h.remove()
for h in handlesd:
h.remove()
if __name__ == '__main__':
import argparse
import os
from datautils import *
from modelutils 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(
'mode', type=str, choices=['nmprune', 'gen', 'load'],
help='Operation mode of the script; "nmprune" for N:M pruning, "gen" for database generation, and "load" for profile evaluation.'
)
parser.add_argument(
'--collect_to', type=str, default='',
help='Folder to store database in; only used in "gen" mode.'
)
parser.add_argument(
'--stitch_from', type=str, default='',
help='Folder to load database from; only used in "load" mode.'
)
parser.add_argument(
'--profile', default='',
help='Profile to load; only used in "load" mode.'
)
parser.add_argument(
'--save', default='',
help='Whether and where to save the resulting checkpoint; not used in "gen" mode.'
)
parser.add_argument(
'--nmblocksize', type=int, default=4,
help='Blocksize for N:M pruning.'
)
parser.add_argument(
'--blocksize', type=int, default=4,
help='Blocksize used for block pruning.'
)
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.'
)
parser.add_argument(
'--min-sparsity', type=float, default=.4,
help='Minimum database sparsity.'
)
parser.add_argument(
'--max-sparsity', type=float, default=.99,
help='Maximum database sparsity.'
)
parser.add_argument(
'--steps', type=int, default=40,
help='Number of equal relative steps between min and max sparsity.'
)
parser.add_argument(
'--batchsize', type=int, default=32,
help='AdaPrune and global AdaPrune batchsize.'
)
parser.add_argument(
'--iters_layerw', type=int, default=10,
help='Number of dataset passes for layer-wise AdaPrune.'
)
parser.add_argument(
'--iters_global', type=int, default=100,
help='Number of dataset passes for global AdaPrune.'
)
parser.add_argument(
'--lr_layerw', type=float, default=1e-3,
help='Learning rate for layer-wise AdaPrune.'
)
parser.add_argument(
'--lr_global', type=float, default=1e-5,
help='Learning rate for global AdaPrune.'
)
args = parser.parse_args()
dataloader, testloader = get_loaders(
args.dataset, path=args.datapath,
nsamples=args.nsamples, seed=args.seed,
batchsize=args.batchsize
)
get_model, test, run = get_functions(args.model)
modelp = get_model()
modeld = get_model()
layersp = find_layers(modelp)
pruners = {}
if args.mode == 'load':
from database import *
db = UnstrDatabase(args.stitch_from, modelp)
db.load_file(modelp, args.profile)
pruners = {}
for name in layersp:
pruners[name] = MagLayerPruner(
layersp[name],
torch.mean((layersp[name].weight == 0).float()).item(),
lr=args.lr_layerw
)
test(modelp, testloader)
if args.iters_global > 0:
global_adaprune(
pruners, modelp, modeld, dataloader, run, iters=args.iters_global, lr=args.lr_global
)
test(modelp, testloader)
if args.save:
torch.save(modelp.state_dict(), args.save)
exit()
if args.mode == 'gen':
if not os.path.exists(args.collect_to):
os.makedirs(args.collect_to)
params = []
for n, p in modelp.named_parameters():
if ('weight' not in n) or (len(p.shape) == 1):
continue
params.append(n.replace('.weight', ''))
modelp = modelp.cpu()
torch.save(modelp.state_dict(), os.path.join(args.collect_to, '0000.pth'))
modelp = modelp.to(DEV)
density = 1 - args.min_sparsity
delta = ((1 - args.max_sparsity) / density) ** (1 / args.steps)
for _ in range(args.steps + 1):
print('%.4f' % (1 - density))
for name in params:
if args.blocksize > 1:
pruners[name] = BlockLayerPruner(layersp[name], args.blocksize, 1 - density, lr=args.lr_layerw)
else:
pruners[name] = MagLayerPruner(layersp[name], 1 - density, lr=args.lr_layerw)
layerw_adaprune(pruners, modeld, dataloader, run, iters=args.iters_layerw)
modelp = modelp.cpu()
torch.save(
modelp.state_dict(),
os.path.join(
args.collect_to, '%s.pth' % ('%.4f' % (1 - density))[2:]
)
)
modelp = modelp.to(DEV)
density *= delta
exit()
if args.mode == 'nmprune':
params = []
for n, p in modelp.named_parameters():
if ('weight' not in n) or (len(p.shape) == 1):
continue
params.append(n.replace('.weight', ''))
params = [p for p in params if p not in firstlast_names(args.model)]
for name in params:
pruners[name] = NM50LayerPruner(layersp[name], args.nmblocksize, lr=args.lr_layerw)
layerw_adaprune(pruners, modeld, dataloader, run, iters=args.iters_layerw)
test(modelp, testloader)
if args.iters_global:
global_adaprune(
pruners, modelp, modeld, dataloader, run, iters=args.iters_global, lr=args.lr_global
)
test(modelp, testloader)
if args.save:
torch.save(modelp.state_dict(), args.save)