tensorflow学习资源-白红宇

tensorflow学习资源

阅读量：6069 次

发布时间：2019-06-20

本文共 1601 字，大约阅读时间需要 5 分钟。

1.TensorFlow-Slim：

是 tensorflow 较新版本的扩充包，可以简化繁杂的网络定义，其中也提供了一些demo:

AlexNet

InceptionV1/V2/V3

OverFeat

ResNet

例如 VGG-16 网络，寥寥数行就可以定义完毕：

                            
def
vgg16(inputs):
  
with slim.arg_scope([slim.conv2d, slim.fully_connected],
                      
activation_fn
=
tf.nn.relu,
                      
weights_initializer
=
tf.truncated_normal_initializer(
0.0
, 
0.01
),
                      
weights_regularizer
=
slim.l2_regularizer(
0.0005
)):
    
net 
=
slim.repeat(inputs, 
2
, slim.conv2d, 
64
, [
3
, 
3
], scope
=
'conv1'
)
    
net 
=
slim.max_pool2d(net, [
2
, 
2
], scope
=
'pool1'
)
    
net 
=
slim.repeat(net, 
2
, slim.conv2d, 
128
, [
3
, 
3
], scope
=
'conv2'
)
    
net 
=
slim.max_pool2d(net, [
2
, 
2
], scope
=
'pool2'
)
    
net 
=
slim.repeat(net, 
3
, slim.conv2d, 
256
, [
3
, 
3
], scope
=
'conv3'
)
    
net 
=
slim.max_pool2d(net, [
2
, 
2
], scope
=
'pool3'
)
    
net 
=
slim.repeat(net, 
3
, slim.conv2d, 
512
, [
3
, 
3
], scope
=
'conv4'
)
    
net 
=
slim.max_pool2d(net, [
2
, 
2
], scope
=
'pool4'
)
    
net 
=
slim.repeat(net, 
3
, slim.conv2d, 
512
, [
3
, 
3
], scope
=
'conv5'
)
    
net 
=
slim.max_pool2d(net, [
2
, 
2
], scope
=
'pool5'
)
    
net 
=
slim.fully_connected(net, 
4096
, scope
=
'fc6'
)
    
net 
=
slim.dropout(net, 
0.5
, scope
=
'dropout6'
)
    
net 
=
slim.fully_connected(net, 
4096
, scope
=
'fc7'
)
    
net 
=
slim.dropout(net, 
0.5
, scope
=
'dropout7'
)
    
net 
=
slim.fully_connected(net, 
1000
, activation_fn
=
None
, scope
=
'fc8'
)
  
return
net
         
       