-
Notifications
You must be signed in to change notification settings - Fork 15
/
legr.py
295 lines (255 loc) · 15 KB
/
legr.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
import os
import time
import torch
import queue
import argparse
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from utils.drivers import train, test, get_dataloader
from model.MobileNetV2 import MobileNetV2, InvertedResidual
from pruner.fp_mbnetv2 import FilterPrunerMBNetV2
from pruner.fp_resnet import FilterPrunerResNet
class LeGR:
def __init__(self, dataset, datapath, model, pruner, rank_type='l2_weight', batch_size=32, lr=1e-3, safeguard=0, global_random_rank=False, lub='', device='cuda'):
self.device = device
self.sample_for_ranking = 1 if rank_type in ['l1_weight', 'l2_weight', 'l2_bn', 'l1_bn', 'l2_bn_param'] else 5000
self.safeguard = safeguard
self.lub = lub
self.lr = lr
self.img_size = 32 if 'CIFAR' in args.dataset else 224
self.batch_size = batch_size
self.rank_type = rank_type
self.train_loader, self.val_loader, self.test_loader = get_dataloader(self.img_size, dataset, datapath, batch_size, args.no_val)
if 'CIFAR100' in dataset:
num_classes = 100
elif 'CIFAR10' in dataset:
num_classes = 10
elif 'ImageNet' in dataset:
num_classes = 1000
elif 'CUB200' in dataset:
num_classes = 200
self.model = model
self.criterion = torch.nn.CrossEntropyLoss()
self.pruner = eval(pruner)(self.model, rank_type, num_classes, safeguard, random=global_random_rank, device=device)
self.model.train()
def learn_ranking_ea(self, name, model_desc, tau_hat, long_ft, target):
name = name
start_t = time.time()
self.pruner.reset()
self.pruner.model.eval()
self.pruner.forward(torch.zeros((1,3,self.img_size,self.img_size), device=self.device))
original_flops = self.pruner.cur_flops
original_size = self.pruner.cur_size
print('Before Pruning, FLOPs: {:.3f}M, Size: {:.3f}M'.format(original_flops/1e6, original_size/1e6))
mean_loss = []
num_layers = len(self.pruner.filter_ranks)
minimum_loss = 10
best_perturbation = None
POPULATIONS = 64
SAMPLES = 16
GENERATIONS = 400
SCALE_SIGMA = 1
MUTATE_PERCENT = 0.1
index_queue = queue.Queue(POPULATIONS)
population_loss = np.zeros(0)
population_data = []
original_dist = self.pruner.filter_ranks.copy()
original_dist_stat = {}
for k in sorted(original_dist):
a = original_dist[k].cpu().numpy()
original_dist_stat[k] = {'mean': np.mean(a), 'std': np.std(a)}
# Initialize Population
for i in range(GENERATIONS):
step_size = 1-(float(i)/(GENERATIONS*1.25))
# Perturn distribution
perturbation = []
if i == POPULATIONS-1:
for k in sorted(self.pruner.filter_ranks.keys()):
perturbation.append((1,0))
elif i < POPULATIONS-1:
for k in sorted(self.pruner.filter_ranks.keys()):
scale = np.exp(float(np.random.normal(0, SCALE_SIGMA)))
shift = float(np.random.normal(0, original_dist_stat[k]['std']))
perturbation.append((scale, shift))
else:
mean_loss.append(np.mean(population_loss))
sampled_idx = np.random.choice(POPULATIONS, SAMPLES)
sampled_loss = population_loss[sampled_idx]
winner_idx_ = np.argmin(sampled_loss)
winner_idx = sampled_idx[winner_idx_]
oldest_index = index_queue.get()
# Mutate winner
base = population_data[winner_idx]
# Perturb distribution
mnum = int(MUTATE_PERCENT * len(self.pruner.filter_ranks))
mutate_candidate = np.random.choice(len(self.pruner.filter_ranks), mnum)
for k in sorted(self.pruner.filter_ranks.keys()):
scale = 1
shift = 0
if k in mutate_candidate:
scale = np.exp(float(np.random.normal(0, SCALE_SIGMA*step_size)))
shift = float(np.random.normal(0, original_dist_stat[k]['std']))
perturbation.append((scale*base[k][0], shift+base[k][1]))
# Given affine transformations, rank and prune
self.pruner.pruning_with_transformations(original_dist, perturbation, target)
# Re-measure the pruned model in terms of FLOPs and size
self.pruner.reset()
self.pruner.model.eval()
self.pruner.forward(torch.zeros((1,3,self.img_size,self.img_size), device=self.device))
cur_flops = self.pruner.cur_flops
cur_size = self.pruner.cur_size
self.pruner.model = self.pruner.model.to(self.device)
print('Density: {:.3f}% ({:.3f}M/{:.3f}M) | FLOPs: {:.3f}% ({:.3f}M/{:.3f}M)'.format(float(cur_size)/original_size*100, cur_size/1e6, original_size/1e6,
float(cur_flops)/original_flops*100, cur_flops/1e6, original_flops/1e6))
print('Fine tuning to recover from pruning iteration.')
optimizer = optim.SGD(self.pruner.model.parameters(), lr=self.lr, momentum=0.9, weight_decay=5e-4)
if tau_hat > 0:
train(self.model, self.train_loader, self.val_loader, optimizer, epochs=1, steps=tau_hat, run_test=False, device=self.device)
acc, loss = test(self.model, self.val_loader, device=self.device, get_loss=True)
if np.mean(loss) < minimum_loss:
minimum_loss = np.mean(loss)
best_perturbation = perturbation
if i < POPULATIONS:
index_queue.put(i)
population_data.append(perturbation)
population_loss = np.append(population_loss, [np.mean(loss)])
else:
index_queue.put(oldest_index)
population_data[oldest_index] = perturbation
population_loss[oldest_index] = np.mean(loss)
# Restore the model back to origin
model = torch.load(model_desc)
if isinstance(model, nn.DataParallel):
model = model.module
model.eval()
model = model.to(self.device)
self.pruner.model = model
self.model = model
self.pruner.reset()
self.pruner.model.eval()
self.pruner.forward(torch.zeros((1,3,self.img_size,self.img_size), device=self.device))
print('Generation {}, Step: {:.2f}, Min Loss: {:.3f}'.format(i, step_size, np.min(population_loss)))
total_t = time.time() - start_t
print('Finished. Use {:.2f} hours. Minimum Loss: {:.3f}'.format(float(total_t) / 3600, minimum_loss))
if not os.path.exists('./log'):
os.makedirs('./log')
np.savetxt(os.path.join('./log', '{}_ea_loss.txt'.format(name)), np.array(mean_loss))
np.savetxt(os.path.join('./log', '{}_ea_min.data'.format(name)), best_perturbation)
# Use the best affine transformation to obtain the resulting model
self.pruner.pruning_with_transformations(original_dist, best_perturbation, target)
if not os.path.exists('./ckpt'):
os.makedirs('./ckpt')
torch.save(self.pruner.model, os.path.join('ckpt', '{}_bestarch_init.pt'.format(name)))
def prune(self, name, model_name, long_ft, target=-1):
test_acc = []
b4ft_test_acc = []
density = []
flops = []
# Get the accuracy before pruning
acc = test(self.model, self.test_loader, device=self.device)
test_acc.append(acc)
b4ft_test_acc.append(acc)
self.pruner.reset()
self.model.eval()
self.pruner.forward(torch.zeros((1,3,self.img_size,self.img_size), device=self.device))
b4prune_size = self.pruner.cur_size
b4prune_flops = self.pruner.cur_flops
density.append(self.pruner.cur_size)
flops.append(self.pruner.cur_flops)
print('Before Pruning, Acc: {:.2f}%, FLOPs: {:.3f}M, Size: {:.3f}M'.format(acc, b4prune_flops/1e6, b4prune_size/1e6))
# If there is learned affine transformation, load it.
if self.lub != '':
perturbation = np.loadtxt(self.lub)
else:
perturbation = np.array([[1., 0.] for _ in range(len(self.pruner.filter_ranks))])
self.pruner.pruning_with_transformations(self.pruner.filter_ranks, perturbation, target)
self.pruner.reset()
self.model.eval()
self.pruner.forward(torch.zeros((1,3,self.img_size,self.img_size), device=self.device))
cur_flops = self.pruner.cur_flops
cur_size = self.pruner.cur_size
density.append(cur_size)
flops.append(cur_flops)
print('Density: {:.3f}% ({:.3f}M/{:.3f}M) | FLOPs: {:.3f}% ({:.3f}M/{:.3f}M)'.format(cur_size/b4prune_size*100, cur_size/1e6, b4prune_size/1e6,
cur_flops/b4prune_flops*100, cur_flops/1e6, b4prune_flops/1e6))
print('Fine tuning to recover from pruning iteration.')
if not os.path.exists('./ckpt'):
os.makedirs('./ckpt')
print('Saving untrained pruned model...')
torch.save(self.pruner.model, os.path.join('ckpt', '{}_init.t7'.format(name)))
acc = test(self.model, self.test_loader, device=self.device)
b4ft_test_acc.append(acc)
if not os.path.exists('./log'):
os.makedirs('./log')
print('Finished. Going to fine tune the model a bit more')
if long_ft > 0:
optimizer = optim.SGD(self.model.parameters(), lr=self.lr, momentum=0.9, weight_decay=5e-4, nesterov=True)
#scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, long_ft)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [int(long_ft*0.3), int(long_ft*0.6), int(long_ft*0.8)], gamma=0.2)
if args.no_val:
train(self.model, self.train_loader, self.test_loader, optimizer, epochs=long_ft, scheduler=scheduler, device=self.device, name=name)
else:
train(self.model, self.train_loader, self.val_loader, optimizer, epochs=long_ft, scheduler=scheduler, device=self.device, name=name)
acc = test(self.model, self.test_loader, device=self.device)
test_acc.append(acc)
else:
acc = test(self.model, self.test_loader, device=self.device)
test_acc.append(acc)
log = np.stack([np.array(b4ft_test_acc), np.array(test_acc), np.array(density), np.array(flops)], axis=1)
np.savetxt(os.path.join('./log', '{}_test_acc.txt'.format(name)), log)
print('Summary')
print('Before Pruning- Accuracy: {:.3f}, Cost: {:.3f}M'.format(test_acc[0], b4prune_flops/1e6))
print('After Pruning- Accuracy: {:.3f}, Cost: {:.3f}M'.format(test_acc[-1], cur_flops/1e6))
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--name", type=str, default='pruned_mbnetv2', help='Name for the experiments, the resulting model and logs will use this')
parser.add_argument("--datapath", type=str, default='./data', help='Path toward the dataset that is used for this experiment')
parser.add_argument("--dataset", type=str, default='torchvision.datasets.CIFAR10', help='The class name of the dataset that is used, please find available classes under the dataset folder')
parser.add_argument("--model", type=str, default='./ckpt/resnet56_cifar10.t7', help='The pre-trained model that pruning starts from')
parser.add_argument("--pruner", type=str, default='FilterPrunerResNet', help='Different network require differnt pruner implementation')
parser.add_argument("--rank_type", type=str, default='l2_weight', help='The ranking criteria for filter pruning')
parser.add_argument("--lub", type=str, default='', help='The affine transformations')
parser.add_argument("--global_random_rank", action='store_true', default=False, help='When this is specified, none of the rank_type matters, it will randomly prune the filters')
parser.add_argument("--tau_hat", type=int, default=0, help='The number of updates before evaluating for fitness (used in EA).')
parser.add_argument("--long_ft", type=int, default=60, help='It specifies how many epochs to fine-tune the network once the pruning is done')
parser.add_argument("--prune_away", type=float, default=90, help='How many percentage of constraints should be pruned away. E.g., 50 means 50% of FLOPs will be pruned away')
parser.add_argument("--safeguard", type=float, default=0, help='A floating point number that represent at least how many percentage of the original number of channel should be preserved. E.g., 0.10 means no matter what ranking, each layer should have at least 10% of the number of original channels.')
parser.add_argument("--batch_size", type=int, default=32, help='Batch size for training.')
parser.add_argument("--min_lub", action='store_true', default=False, help='Use Evolutionary Algorithm to solve latent variable for minimizing Lipschitz upper bound')
parser.add_argument("--uniform_pruning", action='store_true', default=False, help='Use Evolutionary Algorithm to solve latent variable for minimizing Lipschitz upper bound')
parser.add_argument("--no_val", action='store_true', default=False, help='Use full dataset to train (use to compare with prior art in CIFAR-10)')
parser.add_argument("--cpu", action='store_true', default=False, help='Use CPU')
parser.add_argument("--lr", type=float, default=0.001, help='The learning rate for fine-tuning')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = get_args()
print(args)
print('Pruning {}'.format(args.name))
img_size = 32
device = 'cpu' if args.cpu else 'cuda'
prune_till = -1
prune_away = args.prune_away
model = torch.load(args.model)
if isinstance(model, nn.DataParallel):
model = model.module
model = model.to(device)
legr = LeGR(args.dataset, args.datapath, model, args.pruner, args.rank_type, args.batch_size, args.lr, safeguard=args.safeguard, global_random_rank=args.global_random_rank, lub=args.lub, device=device)
if prune_away > 0:
dummy_size = 32 if 'CIFAR' in args.dataset else 224
legr.pruner.reset()
legr.model.eval()
legr.pruner.forward(torch.zeros((1,3,dummy_size, dummy_size), device=device))
b4prune_flops = legr.pruner.cur_flops
prune_till = b4prune_flops * (1-(prune_away)/100.)
print('Pruned untill {:.3f}M'.format(prune_till/1000000.))
if args.uniform_pruning:
ratio = legr.pruner.get_uniform_ratio(prune_till)
legr.pruner.safeguard = ratio
prune_away = 99
if args.min_lub:
legr.learn_ranking_ea(args.name, args.model, args.tau_hat, args.long_ft, (1-(prune_away)/100.))
else:
legr.prune(args.name, args.model, args.long_ft, (1-(prune_away)/100.))