def MnasNet_block_first(network,last_name='',block_name='conv2_',block_n=3,num_out1=16,num_out2=16,downsampling=False,down_method='pooling',use_global_stats='False'):
input_name=last_name
for i in range(1,block_n+1,1):
if i==1 and downsampling==True:
first_stride=2
else:
first_stride=1
#ConvolutionDepthwise
network,last_name=write_prototxt.ConvolutionDepthwise(network,name=block_name+'/dw',bottom_name=last_name,top_name=block_name+'/dw',num_output=num_out1,
bias_term=False,pad=1,kernel_size=3,stride=first_stride,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+'/dw/bn',name_scale=block_name+"/dw/scale",
bottom_name=block_name+'/dw',top_name=block_name+'/dw',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+"/dw/ReLU",bottom_name=block_name+'/dw',top_name=block_name+'/dw')
network,last_name=write_prototxt.Convolution(network,name=block_name+'/sep',bottom_name=last_name,top_name=block_name+'/sep',num_output=num_out2,
bias_term=False,pad=0,kernel_size=1,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+"/sep/bn",name_scale=block_name+"/sep/scale",
bottom_name=block_name+'/sep',top_name=block_name+'/sep',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+"/sep/ReLU",bottom_name=block_name+'/sep',top_name=block_name+'/sep')
return network,last_name
def ResBlock(network,last_name='',block_name='conv2_',block_n=3,num_out=16,downsampling=False,down_method='pooling',use_global_stats='False'):
input_name=last_name
for i in range(1,block_n+1,1):
if i==1 and downsampling==True:
first_stride=2
else:
first_stride=1
network,last_name=write_prototxt.Convolution(network,name=block_name+str(i)+'_0',bottom_name=last_name,top_name=block_name+str(i)+'_0',num_output=num_out,
bias_term=True,pad=1,kernel_size=3,stride=first_stride,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+str(i)+"_bn"+'0',name_scale=block_name+str(i)+"_scale"+'0',
bottom_name=block_name+str(i)+'_0',top_name=block_name+str(i)+'_0',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+"_ReLU"+'0',bottom_name=block_name+str(i)+'_0',top_name=block_name+str(i)+'_0')
network,last_name=write_prototxt.Convolution(network,name=block_name+str(i)+'_1',bottom_name=last_name,top_name=block_name+str(i)+'_1',num_output=num_out,
bias_term=True,pad=1,kernel_size=3,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+str(i)+"bn"+'1',name_scale=block_name+str(i)+"_scale"+'1',
bottom_name=block_name+str(i)+'_1',top_name=block_name+str(i)+'_1',use_global_stats=use_global_stats)
#network,last_name=ReLU(network,name=block_name+str(i)+"ReLU"+'_1',bottom_name=block_name+str(i)+'_1',top_name=block_name+str(i)+'_1')
if i==1 and downsampling==True:
input_name_temp=last_name
if down_method=='conv':
network,last_name=write_prototxt.Convolution(network,name=block_name+str(i)+'_down',bottom_name=input_name,top_name=block_name+str(i)+'_down',num_output=num_out,
bias_term=True,pad=0,kernel_size=1,stride=first_stride,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+str(i)+"_bn"+'_down',name_scale=block_name+str(i)+"_scale"+'_down',
bottom_name=block_name+str(i)+'_down',top_name=block_name+str(i)+'_down',use_global_stats=use_global_stats)
#network,last_name=ReLU(network,name=block_name+str(i)+"_ReLU"+'_down',bottom_name=block_name+str(i)+'_down',top_name=block_name+str(i)+'_down')
else:
network,last_name=write_prototxt.Pooling(network,name=block_name+str(i)+"_Pooling"+'_down',bottom_name=input_name,top_name=block_name+str(i)+"_Pooling"+'_down',pool='MAX',kernel_size=2,stride=2)
input_name=input_name_temp
network,last_name=write_prototxt.Eltwise(network,name=block_name+"Eltwise_"+str(i),bottom_name1=input_name,
bottom_name2=last_name,top_name=block_name+"Eltwise_"+str(i),operation='SUM')
network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+"ReLU"+'_1',bottom_name=last_name,top_name=last_name)
input_name=last_name
return network,last_name
def MnasNet_block_bd11(network,last_name='',block_name='conv2_',block_n=3,num_out0=16,num_out1=16,num_out2=16,downsampling=False,down_method='pooling',use_global_stats='False'):
input_name=last_name
for i in range(1,block_n+1,1):
if i==1 and downsampling==True:
first_stride=2
else:
first_stride=1
network,last_name=write_prototxt.Convolution(network,name=block_name+'0'+'/expand',bottom_name=last_name,top_name=block_name+'0'+'/expand',num_output=num_out1,
bias_term=False,pad=0,kernel_size=1,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+'0'+"/bn/expand",name_scale=block_name+'0'+"/scale/expand",
bottom_name=block_name+'0'+'/expand',top_name=block_name+'0'+'/expand',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+'0'+"ReLU/dw",bottom_name=block_name+'0'+'/expand',top_name=block_name+'0'+'/expand')
#ConvolutionDepthwise
network,last_name=write_prototxt.ConvolutionDepthwise(network,name=block_name+'0'+'/dw',bottom_name=last_name,top_name=block_name+'0'+'/dw',num_output=num_out1,
bias_term=False,pad=1,kernel_size=3,stride=first_stride,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+'0'+'/bn/dw',name_scale=block_name+'0'+"/scale/dw",
bottom_name=block_name+'0'+'/dw',top_name=block_name+'0'+'/dw',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+'0'+"ReLU/dw",bottom_name=block_name+'0'+'/dw',top_name=block_name+'0'+'/dw')
network,last_name=write_prototxt.Convolution(network,name=block_name+'0'+'/linear',bottom_name=last_name,top_name=block_name+'0'+'/linear',num_output=num_out2,
bias_term=False,pad=0,kernel_size=1,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+'0'+"/bn/linear",name_scale=block_name+'0'+"/scale/linear",
bottom_name=block_name+'0'+'/linear',top_name=block_name+'0'+'/linear',use_global_stats=use_global_stats)
#network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+'0'+"ReLU/dw",bottom_name=block_name+'0'+'/sep',top_name=block_name+'0'+'/sep')
if num_out0==num_out2:
network,last_name = write_prototxt.Eltwise(network,name="block_"+block_name+'0',bottom_name1=input_name,bottom_name2=last_name,top_name="block_"+block_name,operation='')
return network,last_name
def MnasNet_block_bdce(network,last_name='',block_name='conv2_',block_n=3,kernel_size=3,num_out1=16,num_out2=16,downsampling=False,down_method='pooling',use_global_stats='False'):
input_name=last_name
for i in range(1,block_n+1,1):
if i==1 and downsampling==True:
first_stride=2
else:
first_stride=1
network,last_name=write_prototxt.Convolution(network,name=block_name+'/sep1',bottom_name=last_name,top_name=block_name+'/sep1',num_output=num_out1,
bias_term=False,pad=0,kernel_size=1,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+"/sep1/bn",name_scale=block_name+"/sep1/scale",
bottom_name=block_name+'/sep1',top_name=block_name+'/sep1',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+"/sep1/ReLU",bottom_name=block_name+'/sep1',top_name=block_name+'/sep1')
#ConvolutionDepthwise
network,last_name=write_prototxt.ConvolutionDepthwise(network,name=block_name+'/dw',bottom_name=last_name,top_name=block_name+'/dw',num_output=num_out1,
bias_term=False,pad=int(kernel_size/2),kernel_size=kernel_size,stride=first_stride,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+'/dw/bn',name_scale=block_name+"/dw/scale",
bottom_name=block_name+'/dw',top_name=block_name+'/dw',use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name=block_name+"/dw/ReLU",bottom_name=block_name+'/dw',top_name=block_name+'/dw')
network,last_name=write_prototxt.Convolution(network,name=block_name+'/sep2',bottom_name=last_name,top_name=block_name+'/sep2',num_output=num_out2,
bias_term=False,pad=0,kernel_size=1,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn=block_name+"/sep2/bn",name_scale=block_name+"/sep2/scale",
bottom_name=block_name+'/sep2',top_name=block_name+'/sep2',use_global_stats=use_global_stats)
if block_name.split("_")[-1]=="1":
network,last_name=write_prototxt.ReLU(network,name=block_name+"/sep2/ReLU",bottom_name=block_name+'/sep2',top_name=block_name+'/sep2')
if block_name.split("_")[-1]!="1":
network,last_name=write_prototxt.Eltwise(network,name=block_name+"/Eltwise"+str(i),bottom_name1=input_name,
bottom_name2=last_name,top_name=block_name+"/Eltwise"+str(i),operation='SUM')
network,last_name=write_prototxt.ReLU(network,name=block_name+"/Eltwise/ReLU",bottom_name=last_name,top_name=last_name)
return network,last_name
def ResNet_cifar( mode='train',root_path='',batch_size=32,block_n=3):
network='name:"ResNet-20"'+'\n'
if mode=='train':
network,last_name=write_prototxt.data(network,name="Data1",mirror=True,scale=0.00390625,crop_size=32,batch_size=batch_size,backend="LMDB",shuffle=True,datasets_path=root_path)
use_global_stats=False
elif mode=='test':
network,last_name=write_prototxt.data(network,name="Data1",mirror=False,scale=0.00390625,crop_size=32,batch_size=batch_size,backend="LMDB",shuffle=False,datasets_path=root_path)
use_global_stats=True
network,last_name=write_prototxt.Convolution(network,name="conv1",bottom_name=last_name,top_name='conv1',num_output=16,
bias_term=True,pad=1,kernel_size=3,stride=1,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn="conv1/bn",name_scale="conv1/scale",
bottom_name="conv1",top_name="conv1",use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name="conv1/ReLU",bottom_name='conv1',top_name='conv1')
#network=Pooling(network,name="Pooling1",bottom_name='conv1',top_name='Pooling1',pool='MAX')
network,last_name=ResBlock(network,last_name=last_name,block_name='conv2_',block_n=block_n,num_out=16,downsampling=False,use_global_stats=use_global_stats)
network,last_name=ResBlock(network,last_name=last_name,block_name='conv3_',block_n=block_n,num_out=32,downsampling=True,down_method='conv',use_global_stats=use_global_stats)
network,last_name=ResBlock(network,last_name=last_name,block_name='conv4_',block_n=block_n,num_out=64,downsampling=True,down_method='conv',use_global_stats=use_global_stats)
#network,last_name=ResBlock(network,last_name=last_name,block_name='conv5_',block_n=block_n,num_out=64,downsampling=True,down_method='conv',use_global_stats=use_global_stats)
network,last_name=write_prototxt.Pooling(network,name="Pooling1",bottom_name=last_name,top_name='Pooling1',pool='AVE',global_pooling=True)
network,last_name=write_prototxt.InnerProduct(network,name="fc1",bottom_name=last_name,top_name='fc1',num_output=10,weight_type='msra',bias_type='constant')
#if mode=='train':
network,last_name=write_prototxt.SoftmaxWithLoss(network,name="Softmax1",bottom_name1='fc1',bottom_name2='label',top_name='Softmax1')
if mode=='test':
network,last_name=write_prototxt.Accuracy(network,name="prob",bottom_name1='fc1',bottom_name2='label',top_name='prob')
#
#
print network
return network
def Net( mode='train',root_path='',batch_size=32):
network='name:"MnasNet"'+'\n'
if mode=='train':
network,last_name=write_prototxt.data(network,name="data",mirror=True,scale=0.00390625,crop_size=224,batch_size=batch_size,backend="LMDB",shuffle=True,datasets_path=root_path)
use_global_stats=False
elif mode=='test':
network,last_name=write_prototxt.data(network,name="data",mirror=False,scale=0.00390625,crop_size=224,batch_size=batch_size,backend="LMDB",shuffle=False,datasets_path=root_path)
use_global_stats=True
elif mode=='deploy':
network='input: "data"'+'\n'
network=network+'input_dim: 1'+'\n'
network=network+'input_dim: 3'+'\n'
network=network+'input_dim: 224'+'\n'
network=network+'input_dim: 224'+'\n'
last_name="data"
use_global_stats=True
network,last_name=write_prototxt.Convolution(network,name="conv1",bottom_name=last_name,top_name='conv1',num_output=32,
bias_term=False,pad=1,kernel_size=3,stride=2,weight_type='msra',bias_type='constant')
network,last_name=write_prototxt.BatchNorm(network,name_bn="conv1/bn",name_scale="conv1/scale",
bottom_name="conv1",top_name="conv1",use_global_stats=use_global_stats)
network,last_name=write_prototxt.ReLU(network,name="conv1/ReLU",bottom_name='conv1',top_name='conv1')
network,last_name=MnasNet_block_first(network,last_name=last_name,block_name='conv2_1',
block_n=1,num_out1=32,num_out2=16,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv3_1',
block_n=1,kernel_size=3,num_out1=48,num_out2=24,downsampling=True,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv3_2',
block_n=1,kernel_size=3,num_out1=72,num_out2=24,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv3_3',
block_n=1,kernel_size=3,num_out1=72,num_out2=24,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv4_1',
block_n=1,kernel_size=5,num_out1=72,num_out2=40,downsampling=True,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv4_2',
block_n=1,kernel_size=5,num_out1=120,num_out2=40,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv4_3',
block_n=1,kernel_size=5,num_out1=120,num_out2=40,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv5_1',
block_n=1,kernel_size=5,num_out1=240,num_out2=80,downsampling=True,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv5_2',
block_n=1,kernel_size=5,num_out1=480,num_out2=80,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv5_3',
block_n=1,kernel_size=5,num_out1=480,num_out2=80,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv6_1',
block_n=1,kernel_size=3,num_out1=480,num_out2=96,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv6_2',
block_n=1,kernel_size=3,num_out1=576,num_out2=96,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv7_1',
block_n=1,kernel_size=5,num_out1=576,num_out2=192,downsampling=True,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv7_2',
block_n=1,kernel_size=5,num_out1=1152,num_out2=192,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv7_3',
block_n=1,kernel_size=5,num_out1=1152,num_out2=192,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv7_4',
block_n=1,kernel_size=5,num_out1=1152,num_out2=192,downsampling=False,use_global_stats=use_global_stats)
network,last_name=MnasNet_block_bdce(network,last_name=last_name,block_name='conv8_1',
block_n=1,kernel_size=3,num_out1=1152,num_out2=320,downsampling=False,use_global_stats=use_global_stats)
network,last_name=write_prototxt.Pooling(network,name="Pooling1",bottom_name=last_name,top_name='Pooling1',pool='AVE',global_pooling=True)
network,last_name=write_prototxt.InnerProduct(network,name="fc1",bottom_name=last_name,top_name='fc1',num_output=2,weight_type='msra',bias_type='constant')
#if mode=='train':
network,last_name=write_prototxt.SoftmaxWithLoss(network,name="Softmax1",bottom_name1='fc1',bottom_name2='label',top_name='Softmax1')
if mode=='test':
network,last_name=write_prototxt.Accuracy(network,name="prob",bottom_name1='fc1',bottom_name2='label',top_name='prob')
#if mode=='deploy':
#network,last_name=write_prototxt.Softmax(network,name="prob",bottom_name='fc1',top_name='prob')
# print network
return network
def Net(mode='train', root_path='', batch_size=32):
network = 'name:"MobileNetV1"' + '\n'
if mode == 'train':
network, last_name = write_prototxt.data(network,
name="Data1",
mirror=True,
crop_size=28,
batch_size=batch_size,
backend="LMDB",
shuffle=True,
datasets_path=root_path)
use_global_stats = False
elif mode == 'test':
network, last_name = write_prototxt.data(network,
name="Data1",
mirror=False,
crop_size=28,
batch_size=batch_size,
backend="LMDB",
shuffle=False,
datasets_path=root_path)
use_global_stats = True
network, last_name = write_prototxt.Convolution(network,
name="conv1",
bottom_name=last_name,
top_name='conv1',
num_output=32,
bias_term=False,
pad=1,
kernel_size=3,
stride=1,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn="conv1_bn",
name_scale="conv1__scale",
bottom_name="conv1",
top_name="conv1",
use_global_stats=use_global_stats)
network, last_name = write_prototxt.ReLU(network,
name="conv1_ReLU",
bottom_name='conv1',
top_name='conv1')
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv2_',
block_n=1,
num_out1=32,
num_out2=64,
downsampling=False,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv3_',
block_n=1,
num_out1=64,
num_out2=128,
downsampling=True,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv4_',
block_n=1,
num_out1=128,
num_out2=128,
downsampling=False,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv5_',
block_n=1,
num_out1=128,
num_out2=256,
downsampling=False,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv6_',
block_n=1,
num_out1=256,
num_out2=512,
downsampling=True,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv7_',
block_n=1,
num_out1=512,
num_out2=512,
downsampling=False,
use_global_stats=use_global_stats)
network, last_name = Mobile_Unit(network,
last_name=last_name,
block_name='conv8_',
block_n=1,
num_out1=512,
num_out2=512,
downsampling=False,
use_global_stats=use_global_stats)
network, last_name = write_prototxt.Pooling(network,
name="Pooling1",
bottom_name=last_name,
top_name='Pooling1',
pool='AVE',
global_pooling=True)
network, last_name = write_prototxt.InnerProduct(network,
name="fc1",
bottom_name=last_name,
top_name='fc1',
num_output=10,
weight_type='xavier',
bias_type='constant')
if mode == 'train':
network, last_name = write_prototxt.SoftmaxWithLoss(
network,
name="Softmax1",
bottom_name1='fc1',
bottom_name2='Data2',
top_name='Softmax1')
if mode == 'test':
network, last_name = write_prototxt.Accuracy(network,
name="prob",
bottom_name1='fc1',
bottom_name2='Data2',
top_name='prob')
#
#
print network
return network
def ResBlock(network,
last_name='',
block_name='conv2_',
block_n=3,
num_out=16,
downsampling=False,
down_method='pooling',
use_global_stats='False',
dropout_ratio=0.5,
first_layer=False):
input_name = last_name
for i in range(1, block_n + 1, 1):
if i == 1 and downsampling == True:
first_stride = 2
else:
first_stride = 1
network, last_name = write_prototxt.Convolution(
network,
name=block_name + str(i) + '_0',
bottom_name=last_name,
top_name=block_name + str(i) + '_0',
num_output=num_out,
bias_term=True,
pad=1,
kernel_size=3,
stride=first_stride,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + str(i) + "_bn" + '0',
name_scale=block_name + str(i) + "_scale" + '0',
bottom_name=block_name + str(i) + '_0',
top_name=block_name + str(i) + '_0',
use_global_stats=use_global_stats)
network, last_name = write_prototxt.ReLU(
network,
name=block_name + str(i) + "_ReLU" + '0',
bottom_name=block_name + str(i) + '_0',
top_name=block_name + str(i) + '_0')
network, last_name = write_prototxt.Dropout(
network,
name=block_name + "Drop" + str(i),
bottom_name1=block_name + str(i) + '_0',
top_name=block_name + str(i) + '_0',
dropout_ratio=dropout_ratio)
network, last_name = write_prototxt.Convolution(
network,
name=block_name + str(i) + '_1',
bottom_name=last_name,
top_name=block_name + str(i) + '_1',
num_output=num_out,
bias_term=True,
pad=1,
kernel_size=3,
stride=1,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + str(i) + "bn" + '1',
name_scale=block_name + str(i) + "_scale" + '1',
bottom_name=block_name + str(i) + '_1',
top_name=block_name + str(i) + '_1',
use_global_stats=use_global_stats)
#network,last_name=ReLU(network,name=block_name+str(i)+"ReLU"+'_1',bottom_name=block_name+str(i)+'_1',top_name=block_name+str(i)+'_1')
origin_name = last_name
network, last_name = write_prototxt.Pooling(
network,
name=block_name + str(i) + "_Pooling",
bottom_name=last_name,
top_name=block_name + str(i) + '_Pooling',
pool='AVE',
global_pooling=True)
num_output_temp = num_out / 16
if num_output_temp < 16: num_output_temp = 16
num_output_temp *= 10 # 需修改
network, last_name = write_prototxt.Convolution(
network,
name=block_name + str(i) + '_2',
bottom_name=last_name,
top_name=block_name + str(i) + '_2',
num_output=num_output_temp,
bias_term=True,
pad=0,
kernel_size=1,
stride=1,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.ReLU(
network,
name=block_name + str(i) + "_ReLU" + '2',
bottom_name=block_name + str(i) + '_2',
top_name=block_name + str(i) + '_2')
network, last_name = write_prototxt.Convolution(
network,
name=block_name + str(i) + '_3',
bottom_name=last_name,
top_name=block_name + str(i) + '_3',
num_output=num_out,
bias_term=True,
pad=0,
kernel_size=1,
stride=1,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.Sigmoid(
network,
name=block_name + str(i) + "_Prob" + '3',
bottom_name=block_name + str(i) + '_3',
top_name=block_name + str(i) + '_3')
SE_name = last_name
if i == 1 and (downsampling == True or first_layer == True):
input_name_temp = last_name
if down_method == 'conv':
network, last_name = write_prototxt.Convolution(
network,
name=block_name + str(i) + '_down',
bottom_name=input_name,
top_name=block_name + str(i) + '_down',
num_output=num_out,
bias_term=True,
pad=0,
kernel_size=1,
stride=first_stride,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + str(i) + "_bn" + '_down',
name_scale=block_name + str(i) + "_scale" + '_down',
bottom_name=block_name + str(i) + '_down',
top_name=block_name + str(i) + '_down',
use_global_stats=use_global_stats)
#network,last_name=ReLU(network,name=block_name+str(i)+"_ReLU"+'_down',bottom_name=block_name+str(i)+'_down',top_name=block_name+str(i)+'_down')
else:
network, last_name = write_prototxt.Pooling(
network,
name=block_name + str(i) + "_Pooling" + '_down',
bottom_name=input_name,
top_name=block_name + str(i) + "_Pooling" + '_down',
pool='MAX',
kernel_size=2,
stride=2)
input_name = last_name
# network,last_name=write_prototxt.Eltwise(network,name=block_name+"Eltwise_"+str(i),bottom_name1=input_name,
# bottom_name2=last_name,top_name=block_name+"Eltwise_"+str(i),operation='SUM')
network, last_name = write_prototxt.Axpy(
network,
name=block_name + "Axpy_" + str(i),
bottom_name1=SE_name,
bottom_name2=origin_name,
bottom_name3=input_name,
top_name=block_name + "Axpy_" + str(i))
network, last_name = write_prototxt.ReLU(network,
name=block_name + str(i) +
"ReLU" + '_1',
bottom_name=last_name,
top_name=last_name)
input_name = last_name
return network, last_name
def Net(mode='train', root_path='', batch_size=32):
network = 'name:"ShuffleNet"' + '\n'
if mode == 'train':
network, last_name = write_prototxt.data(network,
name="Data1",
mirror=True,
crop_size=28,
batch_size=batch_size,
backend="LMDB",
shuffle=True,
datasets_path=root_path)
use_global_stats = False
elif mode == 'test':
network, last_name = write_prototxt.data(network,
name="Data1",
mirror=False,
crop_size=28,
batch_size=batch_size,
backend="LMDB",
shuffle=False,
datasets_path=root_path)
use_global_stats = True
network, last_name = write_prototxt.Convolution(network,
name="conv1",
bottom_name=last_name,
top_name='conv1',
num_output=15,
bias_term=False,
pad=1,
kernel_size=3,
stride=1,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn="conv1_bn",
name_scale="conv1__scale",
bottom_name="conv1",
top_name="conv1",
use_global_stats=use_global_stats)
network, last_name = write_prototxt.ReLU(network,
name="conv1_ReLU",
bottom_name='conv1',
top_name='conv1')
network, last_name = ShuffleNet_Unit0(network,
last_name=last_name,
block_name='resx1_conv',
block_n=1,
num_out0=24,
num_out1=45,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx2_conv',
block_n=1,
num_out0=30,
num_out1=60,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx3_conv',
block_n=1,
num_out0=30,
num_out1=60,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx4_conv',
block_n=1,
num_out0=30,
num_out1=60,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx5_conv',
block_n=1,
num_out0=30,
num_out1=60,
downsampling=True,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx6_conv',
block_n=1,
num_out0=48,
num_out1=120,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx7_conv',
block_n=1,
num_out0=48,
num_out1=120,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx10_conv',
block_n=1,
num_out0=48,
num_out1=120,
downsampling=True,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx11_conv',
block_n=1,
num_out0=60,
num_out1=240,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx12_conv',
block_n=1,
num_out0=60,
num_out1=240,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = ShuffleNet_Unit(network,
last_name=last_name,
block_name='resx13_conv',
block_n=1,
num_out0=60,
num_out1=240,
downsampling=False,
group=3,
use_global_stats=use_global_stats)
network, last_name = write_prototxt.Pooling(network,
name="Pooling1",
bottom_name=last_name,
top_name='Pooling1',
pool='AVE',
global_pooling=True)
network, last_name = write_prototxt.InnerProduct(network,
name="fc1",
bottom_name=last_name,
top_name='fc1',
num_output=10,
weight_type='xavier',
bias_type='constant')
if mode == 'train':
network, last_name = write_prototxt.SoftmaxWithLoss(
network,
name="Softmax1",
bottom_name1='fc1',
bottom_name2='Data2',
top_name='Softmax1')
if mode == 'test':
network, last_name = write_prototxt.Accuracy(network,
name="prob",
bottom_name1='fc1',
bottom_name2='Data2',
top_name='prob')
#
#
print network
return network
def ShuffleNet_Unit(network,
last_name='',
block_name='conv2_',
block_n=3,
num_out0=16,
num_out1=16,
downsampling=False,
group=3,
down_method='pooling',
use_global_stats='False'):
input_name = last_name
for i in range(1, block_n + 1, 1):
if i == 1 and downsampling == True:
first_stride = 2
else:
first_stride = 1
network, last_name = write_prototxt.Convolution(network,
name=block_name + '1',
bottom_name=last_name,
top_name=block_name +
'1',
num_output=num_out0,
bias_term=False,
pad=0,
kernel_size=1,
stride=1,
weight_type='msra',
bias_type='constant',
group=group)
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + '1' + "_bn",
name_scale=block_name + '1' + "_scale",
bottom_name=block_name + '1',
top_name=block_name + '1',
use_global_stats=use_global_stats)
network, last_name = write_prototxt.ReLU(network,
name=block_name + '1' +
"_ReLU",
bottom_name=block_name + '1',
top_name=block_name + '1')
network, last_name = write_prototxt.ShuffleChannel(
network,
name=block_name + "_shuffle",
bottom_name=last_name,
top_name=block_name + "_shuffle",
group=group)
#ConvolutionDepthwise
network, last_name = write_prototxt.ConvolutionDepthwise(
network,
name=block_name + '2',
bottom_name=last_name,
top_name=block_name + '2',
num_output=num_out0,
bias_term=False,
pad=1,
kernel_size=3,
stride=first_stride,
weight_type='msra',
bias_type='constant')
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + '2' + '_bn',
name_scale=block_name + '2' + "_scale",
bottom_name=block_name + '2',
top_name=block_name + '2',
use_global_stats=use_global_stats)
#network,last_name=write_prototxt.ReLU(network,name=block_name+'2'+"_ReLU",bottom_name=block_name+'2',top_name=block_name+'2')
network, last_name = write_prototxt.Convolution(network,
name=block_name + '3',
bottom_name=last_name,
top_name=block_name +
'3',
num_output=num_out1,
bias_term=False,
pad=0,
kernel_size=1,
stride=1,
weight_type='msra',
bias_type='constant',
group=group)
network, last_name = write_prototxt.BatchNorm(
network,
name_bn=block_name + '3' + "_bn",
name_scale=block_name + '3' + "_scale",
bottom_name=block_name + '3',
top_name=block_name + '3',
use_global_stats=use_global_stats)
#network,last_name=write_prototxt.ReLU(network,name=block_name+str(i)+'0'+"ReLU/dw",bottom_name=block_name+'0'+'/sep',top_name=block_name+'0'+'/sep')
if (downsampling == True):
network, input_name = write_prototxt.Pooling(
network,
name=block_name + "_match",
bottom_name=input_name,
top_name=block_name + "_match",
pool='AVE',
kernel_size=3,
stride=2)
network, last_name = write_prototxt.Concat(
network,
name=block_name + '_concat',
bottom_name1=input_name,
bottom_name2=last_name,
top_name=block_name + '_concat')
network, last_name = write_prototxt.ReLU(
network,
name=block_name + '_concat_ReLU',
bottom_name=block_name + '_concat',
top_name=block_name + '_concat')
else:
network, last_name = write_prototxt.Eltwise(
network,
name=block_name + '_elewise',
bottom_name1=input_name,
bottom_name2=last_name,
top_name=block_name + '_elewise',
operation='SUM')
network, last_name = write_prototxt.ReLU(
network,
name=block_name + 'elewise_ReLU',
bottom_name=block_name + '_elewise',
top_name=block_name + '_elewise')
return network, last_name
