def joinBlock(name, opn, n, inputA, inputB, train=True):
trainparam = []
trainparam2 = []
if train:
trainparam = [dict(lr_mult=1, decay_mult= 1),dict(lr_mult=1, decay_mult= 1)]
trainparam2 = [dict(lr_mult=1, decay_mult= 1),dict(lr_mult=1, decay_mult= 1),dict(lr_mult=1, decay_mult= 1)]
else:
trainparam = [dict(lr_mult=0, decay_mult= 0),dict(lr_mult=0, decay_mult= 0)]
trainparam2 = [dict(lr_mult=0, decay_mult= 0),dict(lr_mult=0, decay_mult= 0),dict(lr_mult=0, decay_mult= 0)]
#if global_stat:
#trainparam2 = [dict(lr_mult=0, decay_mult= 0),dict(lr_mult=0, decay_mult= 0),dict(lr_mult=0, decay_mult= 0)]
#TRAINABLE???? TODO
s = list(name)
del s[0]
if s[0] == 'B':
del s[0]
nname = "".join(s)
if upsample == False:
convolution_param = dict(num_output=opn, kernel_size=2, stride=2, pad=0, weight_filler = dict(type='xavier'))
n["upsample"+name] = L.Deconvolution(n[inputA], convolution_param=convolution_param, param=trainparam, name="upsampleparam"+nname)
else:
convolution_param = dict(num_output=opn, kernel_size=2, stride=2, pad=0, weight_filler = dict(type='constant', value=0.0), bias_term=False)
n["upsample"+name] = L.Deconvolution(n[inputA], convolution_param=convolution_param
,param=[dict(lr_mult=0, decay_mult= 0)], name="upsampleparam"+nname)
n["upsampleB"+name] = L.BatchNorm(n["upsample"+name], use_global_stats=global_stat, param=trainparam2, name="upsampleBparam"+nname)#, param=[{"lr_mult":0},{"lr_mult":0},{"lr_mult":0}])
n[inputB] = L.BatchNorm(n[inputB], use_global_stats=global_stat, param=trainparam2, name=nname+"param")#, param=[{"lr_mult":0},{"lr_mult":0},{"lr_mult":0}])
#ACHTUNG BEI BATCHNORM upsampleB genau da drunter !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
n["concat"+name] = L.Concat(n["upsampleB"+name], n[inputB], concat_param=dict(axis=1))
return n, "concat"+name
def Upsample(self, bottom, name='upsample', factor=2, n_out=3, lr_mult=None, decay_mult=None):
def get_kernel_size(f):
return 2 * f - f % 2
def get_pad(f):
return int( np.ceil( (f-1) / 2.) )
up_k = get_kernel_size(factor)
up_s = factor
up_p = get_pad(factor)
conv_params = {'num_output': n_out, 'group': n_out, \
'kernel_size': up_k, 'pad': up_p, 'stride': up_s, \
'weight_filler': {'type': 'bilinear'}, 'bias_term': False}
if lr_mult is not None:
params = []
for lr, decay in zip(lr_mult, decay_mult):
p = self.layer_param.copy()
p['lr_mult'] = lr
p['decay_mult'] = decay
params.append(p)
return L.Deconvolution( bottom, name=name, param=params, convolution_param=conv_params)
else:
return L.Deconvolution( bottom, name=name, convolution_param=conv_params)
def fcn(split, tops):
n = caffe.NetSpec()
n.data, n.label = L.Python(module='nyud_layers',
layer='NYUDSegDataLayer', ntop=2,
param_str=str(dict(nyud_dir='../data/nyud', split=split,
tops=tops, seed=1337)))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr = L.Convolution(n.drop7, num_output=40, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.upscore2 = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=40, kernel_size=4, stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool4 = L.Convolution(n.pool4, num_output=40, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.score_pool4c = crop(n.score_pool4, n.upscore2)
n.fuse_pool4 = L.Eltwise(n.upscore2, n.score_pool4c,
operation=P.Eltwise.SUM)
n.upscore16 = L.Deconvolution(n.fuse_pool4,
convolution_param=dict(num_output=40, kernel_size=32, stride=16,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore16, n.data)
n.loss = L.SoftmaxWithLoss(n.score, n.label,
loss_param=dict(normalize=False, ignore_label=255))
return n.to_proto()
def fcn(split):
n = caffe.NetSpec()
n.data, n.sem, n.geo = L.Python(module='siftflow_layers',
layer='SIFTFlowSegDataLayer', ntop=3,
param_str=str(dict(siftflow_dir='../data/sift-flow',
split=split, seed=1337)))
# the base net
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
n.conv4_1, n.relu4_1 = conv_relu(n.pool3, 512)
n.conv4_2, n.relu4_2 = conv_relu(n.relu4_1, 512)
n.conv4_3, n.relu4_3 = conv_relu(n.relu4_2, 512)
n.pool4 = max_pool(n.relu4_3)
n.conv5_1, n.relu5_1 = conv_relu(n.pool4, 512)
n.conv5_2, n.relu5_2 = conv_relu(n.relu5_1, 512)
n.conv5_3, n.relu5_3 = conv_relu(n.relu5_2, 512)
n.pool5 = max_pool(n.relu5_3)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=7, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr_sem = L.Convolution(n.drop7, num_output=33, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.upscore_sem = L.Deconvolution(n.score_fr_sem,
convolution_param=dict(num_output=33, kernel_size=64, stride=32,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_sem = crop(n.upscore_sem, n.data)
# losses to make score happy (o.w. loss_sem)
n.loss = L.SoftmaxWithLoss(n.score_sem, n.sem,
loss_param=dict(normalize=False, ignore_label=255))
n.score_fr_geo = L.Convolution(n.drop7, num_output=3, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.upscore_geo = L.Deconvolution(n.score_fr_geo,
convolution_param=dict(num_output=3, kernel_size=64, stride=32,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_geo = crop(n.upscore_geo, n.data)
n.loss_geo = L.SoftmaxWithLoss(n.score_geo, n.geo,
loss_param=dict(normalize=False, ignore_label=255))
return n.to_proto()
def fcn(split):
n = caffe.NetSpec()
pydata_params = dict(split=split, mean=(104.00699, 116.66877, 122.67892),
seed=1337)
if split == 'train':
pydata_params['sbdd_dir'] = '../data/sbdd/dataset'
pylayer = 'SBDDSegDataLayer'
else:
pydata_params['voc_dir'] = '../data/pascal/VOC2011'
pylayer = 'VOCSegDataLayer'
n.data, n.label = L.Python(module='voc_layers', layer=pylayer,
ntop=2, param_str=str(pydata_params))
# the base net
n.conv1, n.relu1 = conv_relu(n.data,96,ks=11,stride=4,pad=100)
n.pool1 = max_pool(n.relu1)
n.lrn1 = lrn(n.pool1)
n.conv2, n.relu2 = conv_relu(n.lrn1, 256, ks=5, stride=1, pad=2, group=2)
n.pool2 = max_pool(n.relu2)
n.lrn2 = lrn(n.pool2)
n.conv3, n.relu3 = conv_relu(n.lrn2, 384, ks=3, stride = 1, pad=1)
n.conv4, n.relu4 = conv_relu(n.relu3, 384, ks=3, stride = 1, pad=1, group=2)
n.conv5, n.relu5 = conv_relu(n.relu4, 256, ks=3, stride = 1, pad=1, group=2)
n.pool5 = max_pool(n.relu5)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 4096, ks=6, stride = 1, pad=0)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 4096, ks=1, stride = 1, pad=0)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
n.score_fr = L.Convolution(n.drop7, num_output=21, kernel_size=1, pad=0, stride = 1,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.upscore2 = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=21, kernel_size=5, stride=2,
bias_term=False),
param=[dict(lr_mult=0)])
n.score_pool2 = L.Convolution(n.lrn2, num_output=21, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
n.score_pool2c = crop(n.score_pool2, n.upscore2)
n.fuse_pool2 = L.Eltwise(n.upscore2, n.score_pool2c, operation=P.Eltwise.SUM)
n.upscore16 = L.Deconvolution(n.fuse_pool2,
convolution_param=dict(num_output=21, kernel_size=31, stride=16,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore16, n.data)
n.loss = L.SoftmaxWithLoss(n.score, n.label,
loss_param=dict(normalize=True, ignore_label=255))
return n.to_proto()
def fcn(split, num_classes=None):
n = caffe.NetSpec()
n.data = L.Input(shape=[dict(dim=[1, 3, 500, 500])])
# conv1
n.conv1_1, n.relu1_1 = conv_relu(n.data, 64, pad=100)
n.conv1_2, n.relu1_2 = conv_relu(n.relu1_1, 64)
n.pool1 = max_pool(n.relu1_2)
# conv2
n.conv2_1, n.relu2_1 = conv_relu(n.pool1, 128)
n.conv2_2, n.relu2_2 = conv_relu(n.relu2_1, 128)
n.pool2 = max_pool(n.relu2_2)
# conv3
n.conv3_1, n.relu3_1 = conv_relu(n.pool2, 256)
n.conv3_2, n.relu3_2 = conv_relu(n.relu3_1, 256)
n.conv3_3, n.relu3_3 = conv_relu(n.relu3_2, 256)
n.pool3 = max_pool(n.relu3_3)
# score
n.score_fr = L.Convolution(n.pool3, num_output=num_classes, kernel_size=1, pad=0,
param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)])
# deconv : 8x
n.upscore = L.Deconvolution(n.score_fr,
convolution_param=dict(num_output=num_classes, kernel_size=16,
stride=8, bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore, n.data)
return n.to_proto()
def mynet():
data, label = L.DummyData(
shape=[dict(dim=[8, 1, 28, 28]),
dict(dim=[8, 1, 1, 1])],
transform_param=dict(scale=1. / 255),
ntop=2)
# CAFFE = 1
# MKL2017 = 3
kwargs = {'engine': 3}
conv1 = L.Convolution(data,
kernel_size=[3, 4, 5],
num_output=3,
pad=[1, 2, 3])
bn1 = L.BatchNorm(conv1, **kwargs)
relu1 = L.ReLU(bn1, **kwargs)
convargs = {
'param':
[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=2)],
'convolution_param':
dict(num_output=64,
kernel_size=2,
stride=2,
engine=P.Convolution.CAFFE,
bias_filler=dict(type='constant', value=0),
weight_filler=dict(type='xavier'))
}
deconv1 = L.Deconvolution(relu1, **convargs)
return to_proto(deconv1)
def deconv(netspec,
input_node,
num_output,
scope,
ks=4,
stride=2,
pad=1,
std=0.01):
"""
:param input_node:
:param num_output:
:param ks:
:param stride:
:param pad:
:param std:
:return:
"""
n = netspec
n.node = L.Deconvolution(input_node,
convolution_param=dict(
num_output=num_output,
kernel_size=ks,
stride=stride,
pad=pad,
weight_filler=dict(type='xavier'),
bias_filler=dict(type='constant', value=0)),
param=[dict(lr_mult=0)],
name='%s' % scope)
return n.node
def fpn_block(net): use_relu=False use_bn=False ConvBNLayer(net, 'stage2_tb', 'newC1', use_bn, use_relu, 128, 1, 0, 1) ConvBNLayer(net, 'stage3_tb', 'newC2', use_bn, use_relu, 128, 1, 0, 1) ConvBNLayer(net, 'stage4_tb', 'newC3', use_bn, use_relu, 128, 1, 0, 1) ConvBNLayer(net, 'ext1/fe1_2', 'p4', use_bn, use_relu, 128, 1, 0, 1) net.upP4 = L.Deconvolution(net.p4, convolution_param=dict(kernel_size=4, stride=2, group=128, num_output=128, pad=1, bias_term=False, weight_filler=dict(type='bilinear')), param=[dict(lr_mult=0, decay_mult=0)]) net.p3 = L.Eltwise(net['newC3'], net.upP4) ConvBNLayer(net, 'p3', 'p3_lateral', use_bn, use_relu, 128, 1, 0, 1) net.upP3 = L.Deconvolution(net.p3_lateral, convolution_param=dict(kernel_size=4, stride=2, group=128, num_output=128, pad=1, bias_term=False, weight_filler=dict(type='bilinear')),param=[dict(lr_mult=0, decay_mult=0)]) net.p2 = L.Eltwise(net['newC2'], net.upP3) ConvBNLayer(net, 'p2', 'p2_lateral', use_bn, use_relu, 128, 1, 0, 1) net.upP2 = L.Deconvolution(net.p2_lateral, convolution_param=dict(kernel_size=4, stride=2, group=128, num_output=128, pad=1, bias_term=False, weight_filler=dict(type='bilinear')),param=[dict(lr_mult=0, decay_mult=0)]) net.p1 = L.Eltwise(net['newC1'], net.upP2) return net
def test_nd_conv(self):
"""
ND conv maps the same way in more dimensions.
"""
n = caffe.NetSpec()
# define data with 3 spatial dimensions, otherwise the same net
n.data = L.Input(shape=dict(dim=[2, 3, 100, 100, 100]))
n.conv = L.Convolution(n.data,
num_output=10,
kernel_size=[3, 3, 3],
stride=[1, 1, 1],
pad=[0, 1, 2])
n.pool = L.Pooling(n.conv,
pool=P.Pooling.MAX,
kernel_size=2,
stride=2,
pad=0)
n.deconv = L.Deconvolution(n.pool,
num_output=10,
kernel_size=4,
stride=2,
pad=0)
ax, a, b = coord_map_from_to(n.deconv, n.data)
self.assertEquals(ax, 1)
self.assertTrue(len(a) == len(b))
self.assertTrue(np.all(a == 1))
self.assertEquals(b[0] - 1, b[1])
self.assertEquals(b[1] - 1, b[2])
def upconv(self,
run_shape,
bottom,
num_output_dec,
num_output_conv,
weight_std=0.01,
kernel_size=[2],
stride=[2]):
# Shape update rules
update = RunShapeUpdater()
update.fmaps_update = lambda x: num_output_dec
update.shape_update = lambda x: [
kernel_size[min(i,
len(kernel_size) - 1)] * x[i]
for i in range(0, len(run_shape[-1].shape))
]
self.update_shape(run_shape, update)
deconv = L.Deconvolution(bottom,
convolution_param=dict(
num_output=num_output_dec,
kernel_size=kernel_size,
stride=stride,
pad=[0],
dilation=[1],
group=num_output_dec,
weight_filler=dict(type='constant',
value=1),
bias_term=False),
param=dict(lr_mult=0, decay_mult=0))
# The convolution buffer and weight memory
weight_mem = fsize * num_output_conv * num_output_dec
conv_buff = fsize * run_shape[-1].fmaps
for i in range(0, len(run_shape[-1].shape)):
conv_buff *= 2
conv_buff *= run_shape[-1].shape[i]
# Shape update rules
update.conv_buffer_mem_update = lambda x: conv_buff
update.weight_mem_update = lambda x: weight_mem
update.fmaps_update = lambda x: num_output_conv
update.shape_update = lambda x: [
x[i] for i in range(0, len(run_shape[-1].shape))
]
self.update_shape(run_shape, update)
conv = L.Convolution(deconv,
num_output=num_output_conv,
kernel_size=[1],
stride=[1],
pad=[0],
dilation=[1],
group=1,
param=[dict(lr_mult=1),
dict(lr_mult=2)],
weight_filler=dict(type='gaussian',
std=weight_std),
bias_filler=dict(type='constant'))
return deconv, conv
def upsample(bottom, stride, nout=1, name=None):
s, k, pad = stride, 2 * stride, int(ceil(stride-1)/2)
if not name:
name = "upsample%d"%s
return L.Deconvolution(bottom, name=name, convolution_param=dict(num_output=nout, bias_term=False,
kernel_size=k, stride=s, pad=pad, weight_filler = dict(type="bilinear"), group=nout),
param=[dict(lr_mult=0, decay_mult=0)])
def fcn(split, tops):
n = caffe.NetSpec()
n.color, n.hha, n.label = L.Python(module='nyud_layers',
layer='NYUDSegDataLayer',
ntop=3,
param_str=str(
dict(nyud_dir='../data/nyud',
split=split,
tops=tops,
seed=1337)))
n = modality_fcn(n, 'color', 'color')
n = modality_fcn(n, 'hha', 'hha')
n.score_fused = L.Eltwise(n.score_frcolor,
n.score_frhha,
operation=P.Eltwise.SUM,
coeff=[0.5, 0.5])
n.upscore = L.Deconvolution(n.score_fused,
convolution_param=dict(num_output=40,
kernel_size=64,
stride=32,
bias_term=False),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore, n.color)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
def add_yolo_loss(net,out_layer=0,version='yolov2'):
bottom = net.tops.keys()[-1]
if version is 'yolov2' :
net.conv_global = L.Convolution(net[bottom], kernel_size=1, stride=1,name='conv_global',
num_output=125, pad=0, bias_term=True, weight_filler=dict(type='xavier'), bias_filler=dict(type='constant'))
net.RegionLoss = L.RegionLoss(net.conv_global,net.label,num_class=20,coords=4,num=5,softmax=1,jitter=0.2,rescore=0,object_scale=5.0,noobject_scale=1.0,class_scale=1.0,
coord_scale=1.0,absolute=1,thresh=0.6,random=0,biases=[1.08,1.19,3.42,4.41,6.63,11.38,9.42,5.11,16.62,10.52], include={'phase':caffe.TRAIN})
return net.conv_global,net.conv_global
else :
#net.conv_global1 = L.Convolution(net[bottom], kernel_size=1, stride=1,name='conv_global1',
# num_output=75, pad=0, bias_term=True, weight_filler=dict(type='msra'), bias_filler=dict(type='constant'))
net.Yolov3Loss1 = L.Yolov3(net[bottom],net.label,num_class=20,num=3,object_scale=5.0,noobject_scale=1.0,class_scale=1.0,side=13,
coord_scale=1.0,thresh=0.6,anchors_scale=32,use_logic_gradient=False,mask=[3,4,5],biases=[10,14,23,27,37,58,81,82,135,169,344,319], include={'phase':caffe.TRAIN})
net.upsample = L.Deconvolution(net[bottom],param=dict(lr_mult=0, decay_mult=0),convolution_param=dict(kernel_size=4,stride=2,pad=1,group=75,bias_term=False,num_output=75,weight_filler=dict(type='bilinear')))
x = L.Eltwise(net.upsample, out_layer)
#net.conv_global2 = L.Convolution(x, kernel_size=1, stride=1,name='conv_global2',
# num_output=75, pad=0, bias_term=True, weight_filler=dict(type='msra'), bias_filler=dict(type='constant'))
net.Yolov3Loss2 = L.Yolov3(x,net.label,num_class=20,num=3,object_scale=5.0,noobject_scale=1.0,class_scale=1.0,side=26,
coord_scale=1.0,thresh=0.6,anchors_scale=16,use_logic_gradient=False,mask=[0,1,2],biases=[10,14,23,27,37,58,81,82,135,169,344,319], include={'phase':caffe.TRAIN})
return net[bottom],x
def unet2(bottom, ch, deconv):
def block1(bottom):
return unet_conv(bottom, 256, 128, se=True)
def block2(bottom):
conv1 = unet_conv(bottom, 64, 128, se=True)
ub1 = unet_branch(conv1, block1, 128, 128, 4)
conv2 = unet_conv(ub1, 64, 64, se=True)
return conv2
conv1 = unet_conv(bottom, 32, 64, se=False)
ub1 = unet_branch(conv1, block2, 64, 64, 16)
conv2 = conv_relu(ub1, 64)
if deconv:
return L.Deconvolution(conv2,
convolution_param=dict(num_output=ch,
kernel_size=4,
pad=3,
stride=2))
else:
return L.Convolution(conv2,
kernel_size=3,
stride=1,
num_output=ch,
pad=0)
def deconvolution_layer(net,
inputlayer,
layername,
kernel_size,
stride,
pad,
num_output,
w_lr=1.0,
b_lr=1.0,
init_bias=0.0):
parname_stem = layername
deconv = L.Deconvolution(inputlayer,
convolution_param=dict(
num_output=num_output,
group=num_output,
kernel_size=kernel_size,
stride=stride,
pad=pad,
weight_filler=dict(type="bilinear"),
bias_filler=dict(type="constant",
value=init_bias)),
param=[
dict(name="par_%s_deconv_w" % (parname_stem),
lr_mult=w_lr),
dict(name="par_%s_conv_b" % (parname_stem),
lr_mult=b_lr)
])
net.__setattr__(layername, deconv)
return deconv
def Deconvolution(caffe_net,
layer_idx,
bottom_blob,
out_planes,
kernel_size=2,
stride=2,
padding=0,
skip_connect=None):
names = ['deconv{}'.format(layer_idx)]
caffe_net[names[0]] = L.Deconvolution(bottom_blob,
param={
"lr_mult": 1,
"decay_mult": 1
},
convolution_param=dict(
num_output=out_planes,
bias_term=False,
kernel_size=kernel_size,
stride=stride,
pad=padding))
bottom_blob = caffe_net[names[0]]
if skip_connect is not None:
caffe_net['res{}'.format(layer_idx)] = L.Eltwise(
bottom_blob, skip_connect, operation=P.Eltwise.SUM, coeff=[1, 1])
bottom_blob = caffe_net['res{}'.format(layer_idx)]
return bottom_blob, layer_idx + 1
def upconv(run_shape, bottom, num_output_dec, num_output_conv, weight_std=0.01, kernel_size=[2], stride=[2]):
# Shape update rules
update = [lambda x: 0, lambda x: 0, lambda x: num_output_dec]
update += [[lambda x: x, lambda x: x, lambda x: x]]
update += [[lambda x, i=i: kernel_size[min(i,len(kernel_size)-1)] * x for i in range(0,len(run_shape[-1][4]))]]
update_shape(run_shape, update)
deconv = L.Deconvolution(bottom, convolution_param=dict(num_output=num_output_dec, kernel_size=kernel_size, stride=stride, pad=[0], kstride=[1], group=num_output_dec,
weight_filler=dict(type='constant', value=1), bias_term=False),
param=dict(lr_mult=0, decay_mult=0))
# The convolution buffer and weight memory
weight_mem = fsize * num_output_conv * num_output_dec
conv_buff = fsize * run_shape[-1][2]
for i in range(0,len(run_shape[-1][4])):
conv_buff *= 2
conv_buff *= run_shape[-1][4][i]
# Shape update rules
update = [lambda x: conv_buff, lambda x: weight_mem, lambda x: num_output_conv]
update += [[lambda x: x, lambda x: x, lambda x: x]]
update += [[lambda x, i=i: x for i in range(0,len(run_shape[-1][4]))]]
update_shape(run_shape, update)
conv = L.Convolution(deconv, num_output=num_output_conv, kernel_size=[1], stride=[1], pad=[0], kstride=[1], group=1,
param=[dict(lr_mult=1),dict(lr_mult=2)],
weight_filler=dict(type='gaussian', std=weight_std),
bias_filler=dict(type='constant'))
return deconv, conv
def make_upsample(bottom, num_classes):
return L.Deconvolution(
bottom,
param=[dict(lr_mult=0, decay_mult=0)],
convolution_param=dict(
bias_term=False, num_output=num_classes, kernel_size=16, stride=8,
group=num_classes, pad=4, weight_filler=dict(type="bilinear")))
def upsample_merge(n, data, to_upsample, width, iters):
deconv = L.Deconvolution(to_upsample,
convolution_param=dict(num_output=width,
kernel_size=2,
stride=2,
weight_filler=get_filler(),
bias_filler=dict([
('type', 'constant'),
('value', 0)
])),
param=[{
"lr_mult": 1,
"decay_mult": 1
}, {
"lr_mult": 2,
"decay_mult": 0
}])
prelu = L.PReLU(deconv)
concat = L.Concat(data, prelu)
netset(n, "concat_up_" + str(width), concat)
left = left_branch(concat, width * 2, iters)
netset(n, "left_branch_up_" + str(width), left)
eltwise = L.Eltwise(concat, left, operation=P.Eltwise.SUM)
netset(n, "elementwise_up_" + str(width), eltwise)
return L.PReLU(eltwise, in_place=True)
def vgg16_ssd_seg(source, bbox_seg_data_param, kwargs, use_batchnorm=False, lr_mult=1):
net = caffe.NetSpec()
net.data, net.bbox, net.seg = L.BBoxSegData(name="data", annotated_data_param=bbox_seg_data_param,
data_param=dict(batch_size=8, backend=P.Data.LMDB, source=source),
ntop=3, **kwargs)
net.cls_specific_bbox, net.binary_mask, net.cls = L.SelectBinary(net.bbox, net.seg, random_select=True, num_class=20, ntop=3)
VGGNetBody(net, from_layer='data', fully_conv=True, reduced=True, dilated=True,
dropout=False, pool_mask=True, freeze_all=True)
AddExtraLayers(net, use_batchnorm, lr_mult=0)
# class vector embedding deconvolution net for class-specific semantic segmentation
net.cls_reshape = L.Reshape(net.cls, shape=dict(dim=[0, 0, 1, 1]))
# add top-down deconvolution net
# mbox_source_layers = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2']
AddExtraTopDownLayers(net, use_batchnorm=True, lr_mult=1)
DeVGGNetBody(net, from_layer='deconv6_1', fully_conv=True, reduced=True, dilated=True,
dropout=False, pool_mask=True)
dekwargs = {
'weight_filler': dict(type='xavier'),
'bias_filler': dict(type='constant', value=0)}
deparam = {'param': [dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)]}
net.seg_score = L.Deconvolution(net.derelu1_1, convolution_param=dict(num_output=2, pad=1, kernel_size=3, **dekwargs), **deparam)
net.seg_loss = L.SoftmaxWithLoss(net.seg_score, net.binary_mask, loss_param=dict(ignore_label=255))
return net.to_proto()
def coord_net_spec(ks=3, stride=1, pad=0, pool=2, dstride=2, dpad=0):
"""
Define net spec for simple conv-pool-deconv pattern common to all
coordinate mapping tests.
"""
n = caffe.NetSpec()
n.data = L.Input(shape=dict(dim=[2, 1, 100, 100]))
n.aux = L.Input(shape=dict(dim=[2, 1, 20, 20]))
n.conv = L.Convolution(n.data,
num_output=10,
kernel_size=ks,
stride=stride,
pad=pad)
n.pool = L.Pooling(n.conv,
pool=P.Pooling.MAX,
kernel_size=pool,
stride=pool,
pad=0)
# for upsampling kernel size is 2x stride
try:
deconv_ks = [s * 2 for s in dstride]
except:
deconv_ks = dstride * 2
n.deconv = L.Deconvolution(n.pool,
num_output=10,
kernel_size=deconv_ks,
stride=dstride,
pad=dpad)
return n
def deconv(prev, nout, ks=4, stride=2, bias_term=False):
return layers.Deconvolution(prev,
convolution_param=dict(num_output=nout,
kernel_size=ks,
stride=stride,
bias_term=bias_term),
param=[dict(lr_mult=0)])
def upconv(netconf, bottom, num_output_dec, num_output_conv, kernel_size=[2], stride=[2], stable_mode=False):
# Stable mode is the more numerically stable pathway
if stable_mode:
deconv = L.Deconvolution(bottom, convolution_param=dict(num_output=num_output_dec, kernel_size=kernel_size, stride=stride, pad=[0], group=num_output_dec,
weight_filler=dict(type='constant', value=1), bias_term=False),
param=dict(lr_mult=0, decay_mult=0))
conv = L.Convolution(deconv, num_output=num_output_conv, kernel_size=[1], stride=[1], pad=[0], group=1,
param=[dict(lr_mult=1),dict(lr_mult=2)],
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
return conv
else:
deconv = L.Deconvolution(bottom, convolution_param=dict(num_output=num_output_conv, kernel_size=kernel_size, stride=stride, pad=[0], group=1,
weight_filler=dict(type='msra'), bias_filler=dict(type='constant')),param=[dict(lr_mult=1),dict(lr_mult=2)])
return deconv
def deconv_bn_relu(bottom, num_output, pad=0, kernel_size=3, stride=1):
deconv = L.Deconvolution(bottom,
param=[dict(lr_mult=1, decay_mult=1)],
convolution_param=dict(
num_output=num_output,
pad=pad,
kernel_size=kernel_size,
stride=stride,
weight_filler=dict(type='msra'),
bias_term=0,
engine=engine))
bn = L.BatchNorm(deconv,
in_place=True,
use_global_stats=use_global_stats,
param=[dict(lr_mult=0),
dict(lr_mult=0),
dict(lr_mult=0)])
scale = L.Scale(bn,
in_place=True,
axis=1,
filler=dict(type='constant', value=1),
bias_term=1,
bias_filler=dict(type='constant', value=0))
relu = L.ReLU(scale, in_place=True, engine=engine)
return deconv, bn, scale, relu
def deconv_act(netconf, bottom, num_output, kernel_size=[3], stride=[1], pad=[0], dilation=[1], group=1, activation='relu'):
deconv = L.Deconvolution(bottom, convolution_param=dict(kernel_size=kernel_size, stride=stride, dilation=dilation,
num_output=num_output, pad=pad, group=group,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant')), param=[dict(lr_mult=1),dict(lr_mult=2)])
# Activation
if activation == 'relu':
relu = L.ReLU(deconv, in_place=True, negative_slope=netconf.relu_slope)
last = relu
if activation == 'tanh':
tanh = L.Tanh(deconv, in_place=True)
last = tanh
if activation == 'sigmoid':
sigm = L.Sigmoid(deconv, in_place=True)
last = sigm
if (netconf.dropout > 0):
drop = L.Dropout(last, in_place=True, dropout_ratio=netconf.dropout)
last = drop
if (netconf.use_batchnorm == True):
bnltrain = L.BatchNorm(last, in_place=True, include=[dict(phase=0)],
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=False, moving_average_fraction=netconf.batchnorm_maf))
bnltest = L.BatchNorm(last, in_place=True, include=[dict(phase=1)],
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=True, moving_average_fraction=netconf.batchnorm_maf))
last = {bnltrain, bnltest}
return last
def deconv(input, num_output, kernel_size, stride, pad):
return L.Deconvolution(input,
convolution_param=dict(bias_term=False,
num_output=num_output,
kernel_size=kernel_size,
stride=stride,
pad=pad))
def conv1_autoencoder(split, batch_sz):
n = caffe.NetSpec()
n.data, n.label = L.ImageData(image_data_param=dict(source=split,
batch_size=batch_sz,
new_height=height,
new_width=width,
is_color=False),
ntop=2)
n.silence = L.Silence(n.label, ntop=0)
n.flatdata_i = L.Flatten(n.data)
n.conv1 = conv(n.data, 5, 5, 64, pad=2)
n.bn1 = L.BatchNorm(n.conv1,
use_global_stats=False,
in_place=True,
param=[{
"lr_mult": 0
}, {
"lr_mult": 0
}, {
"lr_mult": 0
}])
n.scale1 = L.Scale(n.bn1, bias_term=True, in_place=True)
n.relu1 = L.ReLU(n.scale1, relu_param=dict(negative_slope=0.1))
n.pool1 = max_pool(n.relu1, 2, stride=2)
n.code = conv(n.pool1, 5, 5, 64, pad=2)
n.upsample1 = L.Deconvolution(n.code,
param=dict(lr_mult=0, decay_mult=0),
convolution_param=dict(
group=64,
num_output=64,
kernel_size=4,
stride=2,
pad=1,
bias_term=False,
weight_filler=dict(type="bilinear")))
n.deconv1 = conv(n.upsample1, 5, 5, 1, pad=2)
n.debn1 = L.BatchNorm(n.deconv1,
use_global_stats=False,
in_place=True,
param=[{
"lr_mult": 0
}, {
"lr_mult": 0
}, {
"lr_mult": 0
}])
n.descale1 = L.Scale(n.debn1, bias_term=True, in_place=True)
n.derelu1 = L.ReLU(n.descale1, relu_param=dict(negative_slope=0.1))
n.flatdata_o = L.Flatten(n.derelu1)
n.loss_s = L.SigmoidCrossEntropyLoss(n.flatdata_o,
n.flatdata_i,
loss_weight=1)
n.loss_e = L.EuclideanLoss(n.flatdata_o, n.flatdata_i, loss_weight=0)
return str(n.to_proto())
def deconv_crop_layer(bottom, bottom2,num_out, size_kerbel,size_stride,num_offset):
deconv1 = L.Deconvolution(bottom,
convolution_param = dict(num_output = num_out, kernel_size = size_kerbel, stride = size_stride),
param = [{'lr_mult':0,'decay_mult':1},{'lr_mult':0, 'decay_mult':0}])
feature_dsn = L.Crop(deconv1,bottom2,
crop_param = dict(axis = 2, offset = num_offset))
return feature_dsn
def generate_net():
n = caffe.NetSpec()
n.data = L.Data(source=TRAIN_LMDB_DATA_FILE,
backend=P.Data.LMDB,
batch_size=1,
ntop=1,
transform_param=dict(scale=1. / 255))
n.label = L.Data(source=TRAIN_LMDB_LABEL_FILE,
backend=P.Data.LMDB,
batch_size=1,
ntop=1)
# the base net
n.conv1, n.relu1 = conv_relu(n.data, 11, 96, stride=4, pad=100)
n.pool1 = max_pool(n.relu1, 3, stride=2)
n.norm1 = L.LRN(n.pool1, local_size=5, alpha=1e-4, beta=0.75)
n.conv2, n.relu2 = conv_relu(n.norm1, 5, 256, pad=2, group=2)
n.pool2 = max_pool(n.relu2, 3, stride=2)
n.norm2 = L.LRN(n.pool2, local_size=5, alpha=1e-4, beta=0.75)
n.conv3, n.relu3 = conv_relu(n.norm2, 3, 384, pad=1)
n.conv4, n.relu4 = conv_relu(n.relu3, 3, 384, pad=1, group=2)
n.conv5, n.relu5 = conv_relu(n.relu4, 3, 256, pad=1, group=2)
n.pool5 = max_pool(n.relu5, 3, stride=2)
# fully conv
n.fc6, n.relu6 = conv_relu(n.pool5, 6, 4096)
n.drop6 = L.Dropout(n.relu6, dropout_ratio=0.5, in_place=True)
n.fc7, n.relu7 = conv_relu(n.drop6, 1, 4096)
n.drop7 = L.Dropout(n.relu7, dropout_ratio=0.5, in_place=True)
# weight_filler=dict(type='gaussian', std=0.0001), bias_filler=dict(type='constant')
n.score_fr_ = L.Convolution(
n.drop7,
num_output=2,
kernel_size=1,
pad=0,
param=[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)])
n.upscore_ = L.Deconvolution(n.score_fr_,
convolution_param=dict(
num_output=2,
kernel_size=63,
stride=32,
group=2,
bias_term=False,
weight_filler=dict(type='bilinear')),
param=[dict(lr_mult=0)])
n.score = crop(n.upscore_, n.data)
n.loss = L.SoftmaxWithLoss(n.score,
n.label,
loss_weight=1,
loss_param=dict(normalize=False,
ignore_label=255))
return n.to_proto()
