-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy path2_extract_features.py
174 lines (123 loc) · 5.35 KB
/
2_extract_features.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
# coding: utf-8
# # 特征抽取
#
# Gluon中的model_zoo中模型的一般结构都是分为 feature 和 classifier部分。
#
# 这里使用gluon中已经Pertrained好的network抽取特征，然后再接下来的分类器部分，我们用来组合特征进行分类。
# In[ ]:
# import packages
import mxnet as mx
from mxnet import image
from mxnet import init
from mxnet import gluon
from mxnet import ndarray as nd
from tqdm import tqdm
from mxnet import autograd
from mxnet.gluon.model_zoo.vision import *
import numpy as np
import time
import h5py
import os
import gc
from time import time
# In[ ]:
# 尝试在gpu上运行，cpu也是可以的，只是比较耗时
ctx = mx.gpu()
# ## 定义图片增广
# In[ ]:
def get_augs(shape):
train_augs = [
image.HorizontalFlipAug(.5),
image.RandomCropAug((shape,shape))
]
valid_augs = train_augs
tests_augs = [
image.CenterCropAug((shape,shape))
]
return train_augs, valid_augs, tests_augs
def transform(data, label, augs):
data = data.astype('float32')
for aug in augs:
data = aug(data)
data = nd.transpose(data, (2,0,1))
return data, nd.array([label]).asscalar().astype('float32')
# In[ ]:
loader = gluon.data.DataLoader
# ## 使用gluon中非常好用的 **ImageFolderDataset** 加载数据，实际上就是一个迭代器
# 通过aug支持reshape成network对应需要的input_shape大小
# In[ ]:
def get_images_data_loader(batch_size, shape):
train_augs, valid_augs, tests_augs = get_augs(shape)
train_imgs = gluon.data.vision.ImageFolderDataset('input/train',transform=lambda X, y: transform(X, y, train_augs))
valid_imgs = gluon.data.vision.ImageFolderDataset('input/valid',transform=lambda X, y: transform(X, y, valid_augs))
tests_imgs = gluon.data.vision.ImageFolderDataset('input/tests',transform=lambda X, y: transform(X, y, tests_augs))
train_data = loader(train_imgs, batch_size, shuffle=True)
valid_data = loader(valid_imgs, batch_size, shuffle=True)
tests_data = loader(tests_imgs, batch_size)
return train_data, valid_data, tests_data
# In[ ]:
batch_size = 64
train_224_data, valid_224_data, tests_224_data = get_images_data_loader(batch_size, 224)
train_299_data, valid_299_data, tests_299_data = get_images_data_loader(batch_size, 299)
# 尝试resnet和densent网络模型
# In[ ]:
# input shape 为224 的网络模型
net_224_list = ['resnet18_v2','resnet34_v2', 'resnet50_v2','resnet101_v1', 'resnet152_v1',
'densenet121','densenet161','densenet169','densenet201']
# input shape 为299 的网络模型
net_299_list = ['inception_v3']
# 定义一个可以通过模型名字获得微调后的模型方法
# In[ ]:
def get_tuning_net(net_name, ctx):
net_function = globals()[net_name] # 动态调用
pretrained_net = net_function(pretrained=True, ctx=ctx)
finetune_net = net_function(classes=120, ctx=ctx)
finetune_net.features = pretrained_net.features
finetune_net.classifier.initialize(init.Xavier())
return finetune_net
# In[ ]:
def get_data_loader(net_input_shape, data_scope):
'''
通过network的输入shape和数据类型（train,vaild,tests）
'''
return globals()[data_scope + "_" + str(net_input_shape) + "_data"]
# 下面定义从pretrained 网络模型中抽取特征的方法，后边可以重复利用
# In[ ]:
def extract_features_by_pretrain_network(net_input_shape, data_scope, rebuild=False):
net_list = globals()["net_%d_list" % net_input_shape]
for name in net_list:
features_file_name = "features/%s_%d_%s_features.h5" % (name, net_input_shape, data_scope)
labels_file_name = "features/%d_%s_labels.h5" % (net_input_shape, data_scope)
if not rebuild and os.path.exists(features_file_name) and os.path.exists(labels_file_name):
print("%s and %s have exsits!" % (features_file_name, labels_file_name))
continue
print("starting extract features for network:%s, data_scope:%s, input_shape:%d" % (name, data_scope, net_input_shape))
net = get_tuning_net(name, ctx)
data_loader = get_data_loader(net_input_shape, data_scope)
features = []
labels = []
for X, y in tqdm(data_loader):
if not os.path.exists(features_file_name):
feature = net.features(X.as_in_context(ctx))
features.append(feature.asnumpy())
if not os.path.exists(labels_file_name):
labels.append(y.asnumpy())
if len(features) > 0:
print("saving features to file: %s" % features_file_name)
features = np.concatenate(features, axis=0)
with h5py.File(features_file_name, "w") as f:
f["features"] = features
if os.path.exists(labels_file_name):
continue
print("saving labels to file: %s" % labels_file_name)
with h5py.File(labels_file_name, "w") as f:
labels = np.concatenate(labels, axis=0)
f["labels"] = labels
# ## 开始抽取各个数据的特征
# In[ ]:
extract_features_by_pretrain_network(224, 'train')
extract_features_by_pretrain_network(224, 'valid')
extract_features_by_pretrain_network(224, 'tests')
#extract_features_by_pretrain_network(299, 'train')
#extract_features_by_pretrain_network(299, 'valid')
#extract_features_by_pretrain_network(299, 'tests')