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
import os
import sys
import logging
import paddle
import argparse
import functools
import math
import time
import numpy as np
sys.path[0] = os.path.join(os.path.dirname("__file__"), os.path.pardir)
from paddleslim.prune import Pruner, save_model
from paddleslim.common import get_logger
from paddleslim.analysis import flops
import models
from utility import add_arguments, print_arguments
import paddle.vision.transforms as T
_logger = get_logger(__name__, level=logging.INFO)
parser = argparse.ArgumentParser(description=__doc__)
add_arg = functools.partial(add_arguments, argparser=parser)
# yapf: disable
add_arg('batch_size', int, 64 * 4, "Minibatch size.")
add_arg('use_gpu', bool, True, "Whether to use GPU or not.")
add_arg('model', str, "MobileNet", "The target model.")
add_arg('pretrained_model', str, "../pretrained_model/MobileNetV1_pretained", "Whether to use pretrained model.")
add_arg('lr', float, 0.1, "The learning rate used to fine-tune pruned model.")
add_arg('lr_strategy', str, "piecewise_decay", "The learning rate decay strategy.")
add_arg('l2_decay', float, 3e-5, "The l2_decay parameter.")
add_arg('momentum_rate', float, 0.9, "The value of momentum_rate.")
add_arg('num_epochs', int, 120, "The number of total epochs.")
add_arg('total_images', int, 1281167, "The number of total training images.")
parser.add_argument('--step_epochs', nargs='+', type=int, default=[30, 60, 90], help="piecewise decay step")
add_arg('config_file', str, None, "The config file for compression with yaml format.")
add_arg('data', str, "mnist", "Which data to use. 'mnist' or 'imagenet'")
add_arg('log_period', int, 10, "Log period in batches.")
add_arg('test_period', int, 10, "Test period in epoches.")
add_arg('model_path', str, "./models", "The path to save model.")
add_arg('pruned_ratio', float, None, "The ratios to be pruned.")
add_arg('criterion', str, "l1_norm", "The prune criterion to be used, support l1_norm and batch_norm_scale.")
add_arg('save_inference', bool, False, "Whether to save inference model.")
# yapf: enable
model_list = models.__all__
def get_pruned_params(args, program):
params = []
if args.model == "MobileNet":
for param in program.global_block().all_parameters():
if "_sep_weights" in param.name:
params.append(param.name)
elif args.model == "MobileNetV2":
for param in program.global_block().all_parameters():
if "linear_weights" in param.name or "expand_weights" in param.name:
params.append(param.name)
elif args.model == "ResNet34":
for param in program.global_block().all_parameters():
if "weights" in param.name and "branch" in param.name:
params.append(param.name)
elif args.model == "PVANet":
for param in program.global_block().all_parameters():
if "conv_weights" in param.name:
params.append(param.name)
return params
def piecewise_decay(args):
step = int(math.ceil(float(args.total_images) / args.batch_size))
bd = [step * e for e in args.step_epochs]
lr = [args.lr * (0.1**i) for i in range(len(bd) + 1)]
learning_rate = paddle.optimizer.lr.PiecewiseDecay(boundaries=bd, values=lr)
optimizer = paddle.optimizer.Momentum(
learning_rate=learning_rate,
momentum=args.momentum_rate,
weight_decay=paddle.regularizer.L2Decay(args.l2_decay))
return optimizer
def cosine_decay(args):
step = int(math.ceil(float(args.total_images) / args.batch_size))
learning_rate = paddle.optimizer.lr.CosineAnnealingDecay(
learning_rate=args.lr, T_max=args.num_epochs)
optimizer = paddle.optimizer.Momentum(
learning_rate=learning_rate,
momentum=args.momentum_rate,
weight_decay=paddle.regularizer.L2Decay(args.l2_decay))
return optimizer
def create_optimizer(args):
if args.lr_strategy == "piecewise_decay":
return piecewise_decay(args)
elif args.lr_strategy == "cosine_decay":
return cosine_decay(args)
def compress(args):
train_reader = None
test_reader = None
if args.data == "mnist":
transform = T.Compose([T.Transpose(), T.Normalize([127.5], [127.5])])
train_dataset = paddle.vision.datasets.MNIST(
mode='train', backend="cv2", transform=transform)
val_dataset = paddle.vision.datasets.MNIST(
mode='test', backend="cv2", transform=transform)
class_dim = 10
image_shape = "1,28,28"
elif args.data == "imagenet":
import imagenet_reader as reader
train_dataset = reader.ImageNetDataset(mode='train')
val_dataset = reader.ImageNetDataset(mode='val')
class_dim = 1000
image_shape = "3,224,224"
else:
raise ValueError("{} is not supported.".format(args.data))
image_shape = [int(m) for m in image_shape.split(",")]
assert args.model in model_list, "{} is not in lists: {}".format(args.model,
model_list)
image = paddle.static.data(
name='image', shape=[None] + image_shape, dtype='float32')
label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
# model definition
model = models.__dict__[args.model]()
out = model.net(input=image, class_dim=class_dim)
avg_cost = paddle.nn.functional.loss.cross_entropy(input=out, label=label)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
val_program = paddle.static.default_main_program().clone(for_test=True)
opt = create_optimizer(args)
opt.minimize(avg_cost)
places = paddle.static.cuda_places(
) if args.use_gpu else paddle.static.cpu_places()
place = places[0]
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
if args.pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(args.pretrained_model, var.name))
_logger.info("Load pretrained model from {}".format(
args.pretrained_model))
paddle.static.load(paddle.static.default_main_program(),
args.pretrained_model, exe)
batch_size_per_card = int(args.batch_size / len(places))
train_loader = paddle.io.DataLoader(
train_dataset,
places=places,
feed_list=[image, label],
drop_last=True,
batch_size=batch_size_per_card,
shuffle=True,
return_list=False,
use_shared_memory=True,
num_workers=16)
valid_loader = paddle.io.DataLoader(
val_dataset,
places=place,
feed_list=[image, label],
drop_last=False,
return_list=False,
use_shared_memory=True,
batch_size=batch_size_per_card,
shuffle=False)
def test(epoch, program):
acc_top1_ns = []
acc_top5_ns = []
for batch_id, data in enumerate(valid_loader):
start_time = time.time()
acc_top1_n, acc_top5_n = exe.run(
program, feed=data, fetch_list=[acc_top1.name, acc_top5.name])
end_time = time.time()
if batch_id % args.log_period == 0:
_logger.info(
"Eval epoch[{}] batch[{}] - acc_top1: {}; acc_top5: {}; time: {}".
format(epoch, batch_id,
np.mean(acc_top1_n),
np.mean(acc_top5_n), end_time - start_time))
acc_top1_ns.append(np.mean(acc_top1_n))
acc_top5_ns.append(np.mean(acc_top5_n))
_logger.info("Final eval epoch[{}] - acc_top1: {}; acc_top5: {}".format(
epoch,
np.mean(np.array(acc_top1_ns)), np.mean(np.array(acc_top5_ns))))
def train(epoch, program):
build_strategy = paddle.static.BuildStrategy()
exec_strategy = paddle.static.ExecutionStrategy()
train_program = paddle.static.CompiledProgram(
program).with_data_parallel(
loss_name=avg_cost.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
for batch_id, data in enumerate(train_loader):
start_time = time.time()
loss_n, acc_top1_n, acc_top5_n = exe.run(
train_program,
feed=data,
fetch_list=[avg_cost.name, acc_top1.name, acc_top5.name])
end_time = time.time()
loss_n = np.mean(loss_n)
acc_top1_n = np.mean(acc_top1_n)
acc_top5_n = np.mean(acc_top5_n)
if batch_id % args.log_period == 0:
_logger.info(
"epoch[{}]-batch[{}] - loss: {}; acc_top1: {}; acc_top5: {}; time: {}".
format(epoch, batch_id, loss_n, acc_top1_n, acc_top5_n,
end_time - start_time))
batch_id += 1
test(0, val_program)
params = get_pruned_params(args, paddle.static.default_main_program())
_logger.info("FLOPs before pruning: {}".format(
flops(paddle.static.default_main_program())))
pruner = Pruner(args.criterion)
pruned_val_program, _, _ = pruner.prune(
val_program,
paddle.static.global_scope(),
params=params,
ratios=[args.pruned_ratio] * len(params),
place=place,
only_graph=True)
pruned_program, _, _ = pruner.prune(
paddle.static.default_main_program(),
paddle.static.global_scope(),
params=params,
ratios=[args.pruned_ratio] * len(params),
place=place)
_logger.info("FLOPs after pruning: {}".format(flops(pruned_program)))
for i in range(args.num_epochs):
train(i, pruned_program)
if i % args.test_period == 0:
test(i, pruned_val_program)
save_model(exe, pruned_val_program,
os.path.join(args.model_path, str(i)))
if args.save_inference:
infer_model_path = os.path.join(args.model_path, "infer_models",
str(i))
paddle.static.save_inference_model(
infer_model_path, [image], [out],
exe,
program=pruned_val_program)
_logger.info("Saved inference model into [{}]".format(
infer_model_path))
def main():
paddle.enable_static()
args = parser.parse_args()
print_arguments(args)
compress(args)
if __name__ == '__main__':
main()