def wrn_expansion(bottom, ks, nout, stride_first_conv=1, pad_first_conv=1):
conv_1_1 = L.Convolution(bottom,
kernel_size=ks,
stride=stride_first_conv,
num_output=nout,
pad=pad_first_conv,
bias_term=False,
weight_filler=dict(type='msra'))
batch_norm = L.BatchNorm(conv_1_1,
use_global_stats=False,
in_place=True,
param=[
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)
])
scale = L.Scale(batch_norm, bias_term=True, in_place=True)
relu = L.ReLU(scale, in_place=True)
conv_2_1 = L.Convolution(relu,
kernel_size=ks,
stride=1,
num_output=nout,
pad=1,
bias_term=False,
weight_filler=dict(type='msra'))
conv_1_2 = L.Convolution(bottom,
kernel_size=1,
stride=stride_first_conv,
num_output=nout,
pad=0,
bias_term=False,
weight_filler=dict(type='msra'))
addition = L.Eltwise(conv_2_1, conv_1_2, operation=P.Eltwise.SUM)
return addition
def convolution_unit(input_layer, k, pad, planes, lr_mult=1, decay_mult=1):
""" Generates a convolution unit (conv + batch_norm + ReLU)
Args:
input_layer: the layer on which to stack the conv unit
k (int): the kernel size
pad (int): the padding size
planes (int): the number of filters
lr_mult (int, optional): the learning rate multiplier (defaults to 1)
decay_mult (int, optional): the weight regularization multiplier
Returns:
obj tuple: the Caffe Layers objects
"""
conv = L.Convolution(input_layer,
kernel_size=k,
pad=pad,
num_output=planes,
weight_filler=dict(type='msra'),
param={'lr_mult': lr_mult, 'decay_mult': decay_mult}
)
bn = L.BatchNorm(conv, in_place=True)
scale = L.Scale(conv, in_place=True, bias_term=True,\
param=[{'lr_mult': lr_mult},{'lr_mult': 2*lr_mult}])
relu = L.ReLU(conv, in_place=True)
return conv, bn, scale, relu
def add_module(bottom, num_output, stride):
conv = L.Convolution(bottom,
param=[{
'lr_mult': 1,
'decay_mult': 0
}],
num_output=num_output,
pad=1,
kernel_size=3,
stride=stride,
bias_term=False,
weight_filler=dict(type='gaussian',
std=round(0.01, 2)))
bn = L.BatchNorm(conv,
moving_average_fraction=round(0.05, 2),
param=[{
'lr_mult': 0
}, {
'lr_mult': 0
}, {
'lr_mult': 0
}],
in_place=True)
scale = L.Scale(conv, bias_term=True, in_place=True)
return conv, bn, scale
def Linear(net,
from_layer,
num_output=1,
kernel_size=1,
stride=1,
pad=0,
num_group=1,
name=None,
suffix=""):
type = "Convolution"
if num_group != 1:
type = "DepthwiseConvolution"
conv = L.Convolution(from_layer,
num_output=num_output,
kernel_size=kernel_size,
group=num_group,
stride=stride,
pad=pad,
bias_term=False,
name="%s%s_conv2d" % (name, suffix),
type=type,
engine=P.Convolution.Engine.CAFFE,
**kwargs)
bn = L.BatchNorm(conv, name="%s%s_batchnormal" % (name, suffix))
return bn
def conv_factory_relu_h_w(bottom,
ks_h,
ks_w,
nout,
stride=1,
pad_h=0,
pad_w=0):
conv = L.Convolution(bottom,
kernel_h=ks_h,
kernel_w=ks_w,
stride=stride,
num_output=nout,
pad_h=pad_h,
pad_w=pad_w,
bias_term=False,
weight_filler=dict(type='msra'))
batch_norm = L.BatchNorm(conv,
in_place=True,
param=[
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)
])
scale = L.Scale(batch_norm, bias_term=True, in_place=True)
relu = L.ReLU(scale, in_place=True)
return relu
def conv_bn(bottom, nout, ks=3, stride=1, pad=0, learn=True):
if learn:
param = [dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)]
else:
param = [dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)]
if isinstance(bottom, str):
conv = L.Convolution(bottom=bottom,
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
param=param,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
else:
conv = L.Convolution(bottom,
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
param=param,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
bn = L.BatchNorm(conv)
# lrn = L.LRN(bn)
return conv, bn
def SingleConv(data,
num_output,
kernel_size=3,
stride=1,
padding=1,
is_train=True):
'''
3D卷积+BatchNorm+Relu
@data：待卷积数据
@num_output：输出通道
@kernel_size：卷积核大小
@stride:步长
@padding:填充
Return：Relu激活后的结果
'''
conv = L.Convolution(data,
kernel_size=kernel_size,
stride=stride,
pad=padding,
num_output=num_output,
weight_filler=dict(type='xavier'))
if is_train:
kwargs = {'engine': 3}
else:
kwargs = {'engine': 3, 'use_global_stats': True}
norm = L.BatchNorm(conv, **kwargs)
scale = L.Scale(norm, bias_term=True)
actv = L.ReLU(scale, engine=3)
return actv
def conv_bn_relu(n, name, top_prev, ks, nout, stride=1, pad=0, loop=1):
for idx in range(loop):
n[str(name) + "_" + str(idx)] = L.Convolution(
top_prev, #name = name,
convolution_param=dict(kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
engine=2,
weight_filler=dict(type='xavier'),
bias_term=False),
param=[dict(lr_mult=1)],
)
top_prev = n[str(name) + "_" + str(idx)]
n[str(name) + '_bn_' + str(idx)] = L.BatchNorm(
top_prev,
batch_norm_param=dict(eps=1e-3, moving_average_fraction=0.99))
top_prev = n[str(name) + '_bn_' + str(idx)]
n[str(name) + '_sc_' + str(idx)] = L.Scale(
top_prev, scale_param=dict(bias_term=True))
top_prev = n[str(name) + '_sc_' + str(idx)]
n[str(name) + '_relu_' + str(idx)] = L.ReLU(top_prev, in_place=True)
top_prev = n[str(name) + '_relu_' + str(idx)]
return top_prev
def conv_bn(name,
input,
output,
kernel_size=3,
stride=1,
pad=1,
activation=True,
dilation=1):
conv = L.Convolution(input,
kernel_size=kernel_size,
stride=stride,
num_output=output,
bias_term=False,
pad=pad,
weight_filler=dict(type='xavier'),
dilation=dilation)
# in-place compute means your input and output has the same memory area,which will be more memory effienct
bn = L.BatchNorm(conv, use_global_stats=True, in_place=True)
# scale = L.Scale(bn,filler=dict(value=1),bias_filler=dict(value=0),bias_term=True, in_place=True)
out = L.Scale(bn, bias_term=True, in_place=True)
if activation is True:
out = L.ReLU(out, in_place=True)
return out
def _block_4in1(major, minor, net, bottom, nout, pad, ks, stride):
branch_flag = '{}_branch{}'.format(major, minor)
conv_layer = 'res{}'.format(branch_flag)
bn_layer = 'bn{}'.format(branch_flag)
scale_layer = 'scale{}'.format(branch_flag)
relu_layer = 'res{}_relu'.format(branch_flag)
if ks == 3: # bottleneck layer, grouped convolutions
net[conv_layer] = L.Convolution(bottom,
num_output=nout,
pad=pad,
kernel_size=ks,
stride=stride,
bias_term=False,
group=32) # Cardinality
else:
net[conv_layer] = L.Convolution(bottom,
num_output=nout,
pad=pad,
kernel_size=ks,
stride=stride,
bias_term=False)
net[bn_layer] = L.BatchNorm(net[conv_layer], in_place=True)
net[scale_layer] = L.Scale(net[bn_layer], bias_term=True, in_place=True)
net[relu_layer] = L.ReLU(net[scale_layer], in_place=True)
return net[relu_layer]
def factorization_conv_mxn(bottom,
num_output=64,
kernel_h=1,
kernel_w=7,
stride=1,
pad_h=3,
pad_w=0):
conv_mxn = L.Convolution(
bottom,
num_output=num_output,
kernel_h=kernel_h,
kernel_w=kernel_w,
stride=stride,
pad_h=pad_h,
pad_w=pad_w,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)],
weight_filler=dict(type='xavier', std=0.01),
bias_filler=dict(type='constant', value=0.2))
conv_mxn_bn = L.BatchNorm(conv_mxn, use_global_stats=False, in_place=True)
conv_mxn_scale = L.Scale(conv_mxn,
scale_param=dict(bias_term=True),
in_place=True)
conv_mxn_relu = L.ReLU(conv_mxn, in_place=True)
return conv_mxn, conv_mxn_bn, conv_mxn_scale, conv_mxn_relu
def conv_relu(bottom, nout, ks=3, stride=2, pad=1, bn=True, test=False):
conv = L.Convolution(
bottom,
param=[dict(lr_mult=1, decay_mult=1)],
convolution_param=dict(
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
bias_term=False,
weight_filler=dict(type="msra"),
engine=1 # DEFAULT = 0; CAFFE = 1; CUDNN = 2;
))
if bn:
conv = L.BatchNorm(
conv,
# param=[dict(lr_mult=0, decay_mult=0),
# dict(lr_mult=0, decay_mult=0),
# dict(lr_mult=0, decay_mult=0)],
batch_norm_param=dict(use_global_stats=test, eps=0.00001),
# inplace=True
)
conv = L.Scale(
conv,
# param=[dict(lr_mult=0, decay_mult=0),
# dict(lr_mult=0, decay_mult=0)
# ],
scale_param=dict(filler=dict(value=1),
bias_term=True,
bias_filler=dict(value=0)),
# inplace=True
)
relu_conv = L.ReLU(conv, in_place=True)
return conv, relu_conv
def conv_BN_scale_relu(bottom,
numout,
kernelsize,
stride,
pad,
bias=False,
groups=1):
conv = L.Convolution(bottom,
kernel_size=kernelsize,
stride=stride,
num_output=numout,
pad=pad,
bias_term=bias,
group=groups,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
BN = L.BatchNorm(conv,
in_place=True,
param=[
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)
])
scale = L.Scale(BN,
in_place=True,
bias_term=True,
filler=dict(value=1),
bias_filler=dict(value=0))
relu = L.ReLU(scale, in_place=True)
return scale, relu
def define_model(self):
n = caffe.NetSpec()
pylayer = 'ClsDataLayer'
pydata_params = dict(phase='train', nodule_cubic=self.nodule_cubic,candidate_cubic=self.candidate_cubic, ratio=self.ratio,
#pydata_params = dict(phase='train', img_root='/workspace/pai/data/',
batch_size=4,crop_size=[40,40,40],random=True)
n.data, n.label = L.Python(module='data.ClsDataLayer', layer=pylayer,
ntop=2, param_str=str(pydata_params))
n.conv1=BasicConv(n.data,32)#(40,40,40)
n.downsample1=Inception_v1(n.conv1,32,32)#(20,20,20)
n.conv2=BasicConv(n.downsample1,64)
n.downsample2=Inception_v1(n.conv2,64,64)#(10,10,10)
n.conv3=BasicConv(n.downsample2,64)
n.downsample3=Inception_v1(n.conv3,64,64)#(5,5,5)
n.conv4=BasicConv(n.downsample3,64)
n.conv5=L.Convolution(n.conv4, kernel_size=3,stride=1,pad=0,num_output=16,weight_filler=dict(type='xavier'))
n.bn5=L.BatchNorm(n.conv5,**kwargs)
n.re5=L.ReLU(n.bn5,**kwargs)
n.fc6=L.InnerProduct(n.re5, num_output=150,weight_filler=dict(type='xavier'))
n.re6=L.ReLU(n.fc6,**kwargs)
n.fc7=L.InnerProduct(n.re6, num_output=2,weight_filler=dict(type='xavier'))
n.loss=L.SoftmaxWithLoss(n.fc7,n.label)
#n.loss=L.Python(module='DiceLoss', layer="DiceLossLayer",
# ntop=1, bottom=[n.probs,n.label])
with open(self.model_def, 'w') as f:
f.write(str(n.to_proto()))
def identity_residual(bottom, kernel_size=3, num_out=64, stride=1, pad=0):
# pre_bn = L.BatchNorm(bottom, in_place=True)
pre_bn = L.BatchNorm(bottom, in_place=False)
pre_scale = L.Scale(pre_bn,
scale_param=dict(bias_term=True),
in_place=True)
pre_relu = L.ReLU(pre_scale, in_place=True)
conv1, bn1, scale1, relu1 = conv_bn_scale_relu(pre_relu,
kernel_size=1,
num_out=num_out,
stride=1,
pad=0)
conv2, bn2, scale2, relu2 = conv_bn_scale_relu(relu1,
kernel_size=3,
num_out=num_out,
stride=stride,
pad=1)
conv3 = L.Convolution(relu2,
kernel_size=1,
num_output=num_out * 4,
stride=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1)],
bias_term=False,
weight_filler=weight_filler)
eltsum = eltsum_block(bottom, conv3)
return pre_bn, pre_scale, pre_relu, \
conv1, bn1, scale1, relu1, \
conv2, bn2, scale2, relu2, \
conv3, eltsum
def conv_factory(bottom, ks, nout, stride=1, pad=0):
conv = L.Convolution(bottom, kernel_size=ks, stride=stride,
num_output=nout, pad=pad, bias_term=False, weight_filler=dict(type='msra'))
batch_norm = L.BatchNorm(conv, in_place=True, batch_norm_param=dict(moving_average_fraction=0.9))
scale = L.Scale(batch_norm, bias_term=True, in_place=True)
relu = L.ReLU(scale, in_place=True)
return relu
def bn_relu_conv(bottom, ks, nout, stride, pad, dropout):
conv1 = L.Convolution(bottom,
kernel_size=1,
stride=1,
num_output=nout,
pad=0,
bias_term=False,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
conv = L.Convolution(conv1,
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
bias_term=False,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
batch_norm = L.BatchNorm(conv,
in_place=True,
param=[
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)
])
scale = L.Scale(batch_norm,
bias_term=True,
in_place=True,
filler=dict(value=1),
bias_filler=dict(value=0))
relu = L.ReLU(scale, in_place=True)
# if dropout>0:
# relu = L.Dropout(relu, dropout_ratio=dropout)
return relu
def conv_factory_relu(bottom, ks, nout, stride=1, pad=0):
conv = L.Convolution(bottom, kernel_size=ks, stride=stride,
num_output=nout, pad=pad, bias_term=True, weight_filler=dict(type='msra'), bias_filler=dict(type='constant'))
batch_norm = L.BatchNorm(conv, in_place=True, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)])
scale = L.Scale(batch_norm, bias_term=True, in_place=True)
relu = L.ReLU(scale, in_place=True)
return relu
def conv_BN_scale_relu(split, bottom, nout, ks, stride,
pad): #对输入作 conv—BN—scale—relu
if bottom == "data":
conv = L.Convolution(bottom="data",
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
bias_term=True,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
else:
conv = L.Convolution(bottom,
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
bias_term=True,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
if split == 'train':
use_global_stats = False #训练的时候我们对 BN 的参数取滑动平均
else:
use_global_stats = True #测试的时候我们直接使用输入的参数
BN = L.BatchNorm(
conv,
batch_norm_param=dict(use_global_stats=use_global_stats),
in_place=True, #BN 的学习率惩罚设置为 0，由 scale 学习
# param = [dict(lr_mult = 0, decay_mult = 0),
# dict(lr_mult = 0, decay_mult = 0),
# dict(lr_mult = 0, decay_mult = 0)]
)
scale = L.Scale(BN, scale_param=dict(bias_term=True), in_place=True)
relu = L.ReLU(scale, in_place=True)
return scale, relu
def preactbottleneck(bottom, ks, nout, stride, pad, groups=1):
batch_norm = L.BatchNorm(bottom, in_place=False, batch_norm_param=dict(use_global_stats=True))
scale = L.Scale(batch_norm, bias_term=True, in_place=True)
relu = L.ReLU(scale, in_place=True)
conv = L.Convolution(relu, kernel_size=ks, stride=stride, group=groups,
num_output=nout, pad=pad, bias_term=False, weight_filler=dict(type='msra'), bias_filler=dict(type='constant'))
return conv
def after_conv(conv):
#in-place compute means your input and output has the same memory area,which will be more memory effienct
bn = L.BatchNorm(conv, use_global_stats=False,in_place=False)
#scale = L.Scale(bn,filler=dict(value=1),bias_filler=dict(value=0),bias_term=True, in_place=True)
scale = L.Scale(bn,bias_term=True, in_place=True)
relu=L.ReLU(scale, in_place=True)
return relu
def convLayer(prev, lrn=False, param_name=None, bn=False, **kwargs):
if param_name:
name1 = param_name + '_kernels'
name2 = param_name + '_bias'
conv = L.Convolution(
prev,
param=[dict(lr_mult=1, name=name1),
dict(lr_mult=2, name=name2)],
weight_filler=dict(type='msra'),
**kwargs)
else:
conv = L.Convolution(prev,
param=[dict(lr_mult=1),
dict(lr_mult=2)],
weight_filler=dict(type='msra'),
**kwargs)
if bn:
bn = L.BatchNorm(conv)
relu = L.ReLU(bn, in_place=True)
else:
relu = L.ReLU(conv, in_place=True)
if lrn:
# optional Local Response Normalization
relu = L.LRN(relu,
lrn_param={
'local_size': min(kwargs['num_output'] / 3, 5),
'alpha': 0.0001,
'beta': 0.75
})
return relu
def bn_relu_conv1(bottom, ks, nout, stride, pad, dropout, idx):
batch_norm = L.BatchNorm(bottom,
in_place=False,
param=[
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)
])
scale = L.Scale(batch_norm,
bias_term=True,
in_place=True,
filler=dict(value=1),
bias_filler=dict(value=0))
relu = L.ReLU(scale, in_place=True)
net['conv-block' + str(idx)] = L.Convolution(
relu,
kernel_size=ks,
stride=stride,
num_output=nout,
pad=pad,
bias_term=False,
weight_filler=dict(type='xavier'),
bias_filler=dict(type='constant'))
if dropout > 0:
conv = L.Dropout(net['conv-block' + str(idx)], dropout_ratio=dropout)
return conv
def add_conv_bn_relu(self, bottom, num_output):
model = self.conv(bottom, num_output, wf=msra_constant)
model = L.BatchNorm(model, use_global_stats=False, in_place=True)
model = L.Scale(model, bias_term=True, in_place=True)
model = L.ReLU(model, in_place=True)
return model
def bn_relu_conv(net, mode, flag, bottom, ks, nout, stride, pad, dropout):
suffix = '{}x{}'.format(ks, ks)
flag_bn = '{}_{}_bn'.format(flag, suffix)
flag_scale = '{}_{}_scale'.format(flag, suffix)
flag_relu = '{}_{}_relu'.format(flag, suffix)
flag_conv = '{}_{}_conv'.format(flag, suffix)
flag_drop = '{}_{}_dropout'.format(flag, suffix)
use_global_stats = False
if mode == 1: # TEST phase
use_global_stats = True
net[flag_bn] = L.BatchNorm(bottom, in_place=False,
batch_norm_param = dict(use_global_stats=use_global_stats),
param=[dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)])
net[flag_scale] = L.Scale(net[flag_bn], bias_term=True, in_place=True,
filler=dict(value=1), bias_filler=dict(value=0))
net[flag_relu] = L.PReLU(net[flag_scale], in_place=True)
net[flag_conv] = L.Convolution(net[flag_relu], num_output=nout,
kernel_size=ks, stride=stride, pad=pad,
weight_filler=dict(type='msra'),
bias_term=False)
if dropout > 0:
net[flag_drop] = L.Dropout(net[flag_conv], dropout_ratio=dropout)
return net[flag_drop]
return net[flag_conv]
def conv_bn_scale_relu(bottom, kernel_size=3, num_out=64, stride=1, pad=0, params=conv_params):
weight_filler = params[0]
bias_filler = params[1]
conv = L.Convolution(bottom, kernel_size=kernel_size, stride=stride, num_output=num_out,
pad=pad, param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
weight_filler=weight_filler, bias_filler=bias_filler)
bn_train = L.BatchNorm(conv, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)],
use_global_stats=False, in_place=True, include=dict(phase=0))
bn_test = L.BatchNorm(conv, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0),
dict(lr_mult=0, decay_mult=0)],
use_global_stats=True, in_place=True, include=dict(phase=1))
scale = L.Scale(conv, scale_param=dict(bias_term=True), in_place=True)
relu = L.ReLU(conv, in_place=True)
return conv, bn_train, bn_test, scale, relu
def batch_norm_net():
n = caffe.NetSpec()
n.data = L.DummyData(dummy_data_param=dict(num=64,
channels=1,
height=28,
width=28,
data_filler=dict(
type="gaussian")))
n.label = L.DummyData(dummy_data_param=dict(num=64,
channels=1,
height=1,
width=1,
data_filler=dict(
type="gaussian")))
n.conv1 = L.Convolution(n.data,
kernel_size=7,
stride=2,
num_output=32,
pad=3)
n.pool1 = L.Pooling(n.conv1, pool=P.Pooling.MAX, kernel_size=2, stride=2)
n.relu1 = L.ReLU(n.pool1, in_place=True)
n.norm1 = L.BatchNorm(n.relu1, moving_average_fraction=0.9, in_place=True)
n.scale1 = L.Scale(n.norm1, bias_term=True, in_place=True)
n.ip2 = L.InnerProduct(n.scale1,
num_output=10,
weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
return n.to_proto()
def conv_bn_scale(bottom,
kernel_size=3,
num_out=64,
stride=1,
pad=0,
params=conv_params):
weight_filler = params[0]
bias_filler = params[1]
conv = L.Convolution(bottom,
kernel_size=kernel_size,
stride=stride,
num_output=num_out,
bias_term=False,
pad=pad,
param=[dict(lr_mult=1, decay_mult=1)],
weight_filler=weight_filler)
bn = L.BatchNorm(conv, in_place=True)
scale = L.Scale(
conv,
scale_param=dict(bias_term=True),
in_place=True,
param=[dict(lr_mult=1, decay_mult=0),
dict(lr_mult=1, decay_mult=0)])
return conv, bn, scale
def add_batchnorm(self, bottom):
bn = cl.BatchNorm(bottom, in_place=True)
bn = cl.Scale(bn,
bias_term=True,
bias_filler=dict(value=0),
in_place=True)
return bn
def BR(caffe_net, layer_idx, bottom_blob):
names = ['bn{}'.format(layer_idx), 'prelu{}'.format(layer_idx)]
caffe_net[names[0]] = L.BatchNorm(bottom_blob)
caffe_net[names[1]] = L.PReLU(caffe_net[names[0]])
return caffe_net[names[1]], layer_idx + 1