-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathnni_pruning.py
265 lines (246 loc) · 11.9 KB
/
nni_pruning.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
# -*-coding: utf-8 -*-
"""
@Author : panjq
@E-mail : [email protected]
@Date : 2021-12-09 19:16:19
"""
import os
import copy
import torch
import torch.nn as nn
import torch.onnx
import torch.nn.functional as F
from nni.compression.pytorch.utils.counter import count_flops_params
from nni.compression.pytorch.speedup import ModelSpeedup
from nni.algorithms.compression.pytorch import pruning
from nni.compression.pytorch import apply_compression_results
def model_pruning(model: nn.Module,
input_size=[1, 3, 128, 128],
sparsity=0.2,
prune_mod="L1",
reuse=False,
speedup=True,
output_prune="pruning_output",
mask_file="",
dependency_aware=True,
device="cpu",
verbose=False,
**kwargs):
"""
使用NNI进行模型剪枝和压缩
https://github.com/microsoft/nni/blob/master/docs/en_US/Compression/compression_pipeline_example.ipynb
use l1filter pruner to prune the model
Note that if you use a compressor that need you to pass a optimizer,
you need a new optimizer instead of you have used above, because NNI might modify the optimizer.
And of course this modified optimizer can not be used in finetuning.
Usage:
model = build_model()
model = model_pruning(model,input_size,sparsity=0.2)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
for epoch in range(0, epochs):
trainer(model, optimizer, criterion, epoch)
evaluator(model)
torch.save(model.state_dict(), "model_pruning.pth")
:param model: Pytorch模型
:param input_size: 模型输入的维度[batch-size,channel,H,W]
:param sparsity: 模型剪枝目标稀疏度,值越大,模型越稀疏，计算量越小；0.5表示剪除50%
:param reuse: 是否复现模型剪枝的结果
False: 进行模型剪枝(pruning+ModelSpeedup),会生成mask文件(mask_naive_l1filter.pth)
True : 复现模型剪枝的结果,需要提供mask_file文件
:param speedup: 是否加速模型
:param output_prune: 模型剪枝输出文件
:param mask_file: reuse=True需要提供模型剪枝的mask文件,默认保存在output_prune目录下(mask_naive_l1filter.pth)
:param dependency_aware 依赖感知模式 https://nni.readthedocs.io/zh/stable/Compression/DependencyAware.html
:param device:
:param config: 模型剪枝配置,用于指定需要剪枝网络层；
如果不指定op_names,默认对所以层进行剪枝
config_list = [{'sparsity': sparsity,'op_types': ['Conv2d'],'op_names': ['conv1']}]
:return:
"""
info = ""
model = model.to(device)
if not os.path.exists(output_prune): os.makedirs(output_prune)
prune_file = os.path.join(output_prune, 'pruned_naive_{}filter.pth'.format(prune_mod))
onnx_file = os.path.join(output_prune, 'pruned_naive_{}filter.onnx'.format(prune_mod))
mask_file = os.path.join(output_prune, 'mask_naive_{}filter.pth'.format(prune_mod)) if not mask_file else mask_file
dummy_input = torch.randn(input_size).to(device)
# 原始模型的计算量和参数量
flops, params, _ = count_flops_params(model, dummy_input, verbose=verbose)
info += f"origin-Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M\n"
# 模型剪枝,会生成mask文件(mask_naive_l1filter.pth)
if not reuse:
"""
Choose a pruner and pruning
use l1filter pruner to prune the model
Note that if you use a compressor that need you to pass a optimizer,
you need a new optimizer instead of you have used above, because NNI might modify the optimizer.
And of course this modified optimizer can not be used in finetuning.
"""
if prune_mod.lower() == "Level".lower():
config = [{'sparsity': sparsity, 'op_types': ['Conv2d']}]
pruner = pruning.LevelPruner(model, config)
elif prune_mod.lower() == "L1".lower():
# op_types : Only Conv2d is supported in L1FilterPruner.
# config = [{'sparsity': sparsity, 'op_types': ['Conv2d'], "exclude": False}]
config = [{'sparsity': sparsity, 'op_types': ['Conv2d']}]
pruner = pruning.L1FilterPruner(model, config, dependency_aware, dummy_input=dummy_input)
elif prune_mod.lower() == "L2".lower():
# op_types : Only Conv2d is supported in L2FilterPruner.
config = [{'sparsity': sparsity, 'op_types': ['Conv2d']}]
pruner = pruning.L2FilterPruner(model, config, dependency_aware, dummy_input=dummy_input)
elif prune_mod.lower() == "FPGM".lower():
# op_types : Only Conv2d is supported in FPGM Pruner
config = [{'sparsity': sparsity, 'op_types': ['Conv2d']}]
pruner = pruning.FPGMPruner(model, config, dependency_aware, dummy_input=dummy_input)
elif prune_mod.lower() == "Slim".lower():
# op_types : Only BatchNorm2d is supported in Slim Pruner.
config = [{'sparsity': sparsity, 'op_types': ['BatchNorm2d']}]
pruner = pruning.SlimPruner(model,
config,
optimizer=None,
trainer=None,
criterion=None,
sparsifying_training_epochs=10)
elif prune_mod.lower() == "Slim".lower():
config = [{'sparsity': sparsity, 'op_types': ['BatchNorm2d']}]
pruner = pruning.ActivationMeanRankFilterPruner()
else:
raise Exception("Error prune_mod:{}".format(prune_mod))
# compress the model, the mask will be updated.
pruner.compress()
# pruner.get_pruned_weights()
# use a dummy input to apply the sparsify.
out = model(dummy_input)
# 剪枝后模型的计算量和参数量
flops, params, _ = count_flops_params(model, dummy_input, verbose=verbose)
info += f"pruner-Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M\n"
# export the sparsified and mask model
pruner.export_model(model_path=prune_file, mask_path=mask_file,
onnx_path=onnx_file, input_shape=dummy_input.shape,
device=device,
opset_version=11)
# speedup the model with provided weight mask.If you use a wrapped model, don't forget to unwrap it.
pruner._unwrap_model()
# 将掩码应用到模型,模型会变得更小，推理延迟也会减小
# apply_compression_results(model, mask_file, device)
if speedup:
if not os.path.exists(mask_file): raise Exception("not found mask file:{}".format(mask_file))
print("load mask file to speed up:{}".format(mask_file))
speed_up = ModelSpeedup(model, dummy_input=dummy_input, masks_file=mask_file)
speed_up.speedup_model()
out = model(dummy_input)
# speedup后模型的计算量和参数量
flops, params, _ = count_flops_params(model, dummy_input, verbose=verbose)
info += f"speedup-Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M\n"
# finetune the model to recover the accuracy.
model = model.to(device)
print(info)
return model
def nni_model_pruning_test(model, input_size=[1, 3, 416, 416], sparsity=0.5, output="output", device="cpu"):
"""
https://zhuanlan.zhihu.com/p/382638682
:param model:
:param input_size:
:param sparsity:
:param output:
:param device:
:return:
"""
from ptflops.flops_counter import get_model_complexity_info
if not os.path.exists(output):
os.makedirs(output)
config = [{
'sparsity': sparsity,
'op_types': ['Conv2d']
# 'op_types': ['BatchNorm2d']
}]
dummy_input = torch.randn(input_size).to(device)
#
origin_path = os.path.join(output, 'origin.pth')
onnx_path = os.path.join(output, 'origin.onnx')
torch.save(model.state_dict(), origin_path)
torch.onnx.export(model,
dummy_input,
onnx_path,
do_constant_folding=True,
verbose=False,
export_params=True,
opset_version=11,
input_names=['input'],
output_names=['output'])
tmp_model = copy.deepcopy(model).to(device)
out_tensor = tmp_model(dummy_input)
flops, params, _ = count_flops_params(tmp_model, dummy_input, verbose=True)
print(f"Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M")
pruner = pruning.FPGMPruner(tmp_model, config)
pruner.compress()
flops, params, _ = count_flops_params(tmp_model, dummy_input, verbose=True)
print(f"Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M")
pruned_model_path = os.path.join(output, 'slim_pruned.pth')
pruned_model_mask = os.path.join(output, 'slim_pruned_mask.pth')
pruned_model_onnx = os.path.join(output, 'slim_pruned.onnx')
pruner.export_model(model_path=pruned_model_path, mask_path=pruned_model_mask,
onnx_path=pruned_model_onnx, input_shape=dummy_input.shape,
device=device,
opset_version=11)
tmp_model = copy.deepcopy(model).to(device)
print('model pruned done.')
# apply_compression_results
apply_compression_results(tmp_model, masks_file=pruned_model_mask)
print('apply_compression_results')
out_tensor = tmp_model(dummy_input)
# Speedup
m_speedup = ModelSpeedup(tmp_model, dummy_input, masks_file=pruned_model_mask)
m_speedup.speedup_model()
out_tensor = tmp_model(dummy_input)
print('speedup_model ')
# # print(model)
slim_speedup_path = os.path.join(output, 'slim_speedup_model.pth')
slim_speedup_onnx = os.path.join(output, 'slim_speedup_model.onnx')
torch.save(tmp_model.state_dict(), slim_speedup_path)
torch.onnx.export(tmp_model, dummy_input, slim_speedup_onnx, verbose=False, opset_version=11)
print('pruned model exported.')
flops, params, _ = count_flops_params(tmp_model, dummy_input, verbose=True)
print(f"Model FLOPs {flops / 1e6:.2f}M, Params {params / 1e6:.2f}M")
class SimpleModel(nn.Module):
def __init__(self):
super(SimpleModel, self).__init__()
self.conv1 = nn.Conv2d(3, 32, 3)
self.conv2 = nn.Conv2d(32, 64, 3)
self.conv3 = nn.Conv2d(64, 128, 3)
self.fc = nn.Linear(128, 256)
self.classifier = nn.Linear(256, 10)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = nn.functional.adaptive_avg_pool2d(x, 1).reshape(x.shape[0], -1)
x = self.fc(x)
x = self.classifier(x)
return x
if __name__ == "__main__":
from torchvision.models.resnet import resnet50, resnet18
from torchvision.models.squeezenet import SqueezeNet
from torchvision.models.mobilenet import MobileNetV2
from segment.models import build_model
from libs.performance import performance
device = "cuda:0"
batch_size = 2
width_mult = 1.0
num_classes = 20
input_size = [1, 3, 224, 224]
net_type = 'modnet_v2'
model = resnet18(pretrained=True)
# model = SimpleModel()
# model = MobileNetV2()
# model = build_model.get_models(net_type, input_size, num_classes, width_mult=width_mult, pretrained=False)
model.eval()
inputs = torch.randn(input_size)
model = model.to((device))
inputs = inputs.to((device))
out = model(inputs)
performance.model_performance(model, inputs)
prune_model = copy.deepcopy(model)
prune_model = model_pruning(prune_model, input_size=input_size, sparsity=0.9, dependency_aware=True, device=device)
performance.model_performance(model, inputs)
performance.model_performance(prune_model, inputs)