def __init__(self, indepth, branch=1, active='relu', nclass=1000, method='split'):
super(SummaryBlock, self).__init__()
assert len(indepth) == branch, '各分类分支的通道数必须全部给定，so， len of <indepth> == branch.'
assert method in self.METHOD, 'Unknown <method> %s for SummaryBlock.' % method
self.indepth = indepth
self.branch = branch
self.active = active
self.nclass = nclass
self.method = method
self.active_fc = True
if method == 'split':
for i in range(branch):
fc_layer = nn.Sequential(
nn.BatchNorm2d(indepth[i]),
Activate(active),
AdaAvgPool(),
ViewLayer(),
nn.Linear(indepth[i], nclass))
setattr(self, 'classifier%s' % (i + 1), fc_layer)
elif method == 'merge':
for i in range(branch):
view_layer = nn.Sequential(
nn.BatchNorm2d(indepth[i]),
Activate(active),
AdaAvgPool(),
ViewLayer())
setattr(self, 'pool_view%s' % (i + 1), view_layer)
self.classifier = nn.Linear(sum(indepth), nclass)
else:
raise NotImplementedError
def __init__(self,
indepth,
branch,
active='relu',
pool='avg',
nclass=1000,
method='split'):
super(RockSummary, self).__init__()
assert len(
indepth) == branch, '各输出分支的通道数必须全部给定，len of <indepth> == branch.'
assert method in self.METHOD, 'Unknown <methond> %s.' % method
self.indepth = indepth
self.branch = branch
self.active_fc = True
self.nclass = nclass
self.method = method
if method == 'split':
for b in range(1, branch + 1):
layer = nn.Sequential(nn.BatchNorm2d(indepth[b - 1]),
Activate(active), AdaAvgPool(),
ViewLayer(),
nn.Linear(indepth[b - 1], nclass))
setattr(self, 'classifier%s' % b, layer)
elif method == 'merge':
for b in range(1, branch + 1):
layer = nn.Sequential(nn.BatchNorm2d(indepth[b - 1]),
Activate(active), AdaAvgPool(),
ViewLayer())
setattr(self, 'pool_view%s' % b, layer)
self.classifier = nn.Linear(sum(indepth), nclass)
else:
raise NotImplementedError
def __init__(self, indepth, growth, active='relu', first=False, after=True, down=False,
trans='A', reduce=0.5, convkp='T', inmode='nearest', classify=0, nclass=1000,
last_branch=1, last_down=False, last_expand=0):
super(SingleCouple, self).__init__()
assert last_branch <= 2, '<last_branch> of SingleCouple should be <= 2...'
self.indepth = indepth
self.growth = growth
self.active = getattr(nn.functional, active)
self.first = first
self.after = after
self.down = down
self.trans = trans
self.reduce = reduce
self.trans_func = self._trans[trans]
self.classify = classify
self.nclass = nclass
self.last_branch = last_branch
self.last_down = last_down
self.last_expand = last_expand
self.inmode = inmode
self.convkp = convkp
kp = {'T': (3, 1), 'O': (1, 0)}[convkp]
first_outdepth = indepth + growth if self.first else growth
self.bn1 = nn.BatchNorm2d(indepth)
self.conv1 = nn.Conv2d(indepth, first_outdepth, 3, stride=2, padding=1, bias=False, dilation=1)
self.bn2 = nn.BatchNorm2d(indepth + growth)
self.conv2 = nn.Conv2d(indepth + growth, growth, kp[0], stride=1, padding=kp[1], bias=False, dilation=1)
if self.classify > 0:
self.classifier = nn.Sequential(
nn.BatchNorm2d(indepth + growth),
Activate(active),
AdaAvgPool(),
ViewLayer(dim=-1),
nn.Linear(indepth + growth, nclass)
)
if self.after:
if self.down:
outdepth = int(math.floor((indepth + growth) * reduce))
self.down_res2 = self.trans_func(indepth + growth, outdepth)
self.down_res1 = self.trans_func(indepth + growth, outdepth)
else:
if self.last_down:
outdepth = indepth + growth + last_expand
if self.last_branch >= 1:
self.down_last2 = self.trans_func(indepth + growth, outdepth)
if self.last_branch >= 2:
self.down_last1 = self.trans_func(indepth + growth, outdepth)
else:
if self.classify > 0 and self.last_branch == 2:
# 此时，最后一个Couple的中间层被当做branch输出而对接在Summary上.
# 因此，删除此Couple自带的Classifier,以免与Summary中的Classifier重复.
delattr(self, 'classifier')
self.classify = 0
print('Note: 1 xfc will be deleted because of duplicate with the last-fc!')
def __init__(self,
indepth,
branch=1,
active='relu',
pool='avg',
nclass=1000):
super(RockSummary, self).__init__()
self.indepth = indepth
self.branch = branch
self.active_fc = True
self.pool = pool
for b in range(1, branch + 1):
layer = nn.Sequential(nn.BatchNorm2d(indepth), Activate(active),
AdaAvgPool(), ViewLayer(),
nn.Linear(indepth, nclass))
setattr(self, 'classifier%s' % b, layer)
def __init__(self, indepth, growth, bottle=4, active='relu', first=False, after=True, down=False, trans='A',
reduce=0.5, classify=0, nclass=1000, last_branch=1, last_down=False, last_expand=0):
super(DoubleCouple, self).__init__()
assert last_branch <= 4, '<last_branch> of DoubleCouple should be <= 4...'
self.indepth = indepth
self.growth = growth
self.bottle = bottle
self.classify = classify
self.nclass = nclass
self.active = getattr(nn.functional, active)
self.first = first
self.after = after
self.down = down
self.trans = trans
self.reduce = reduce
self.trans_func = self._trans[trans]
self.last_branch = last_branch
self.last_down = last_down
self.last_expand = last_expand
if bottle > 0:
interdepth = int(bottle * growth)
assert interdepth == math.ceil(bottle * growth), \
'<bottle> * <growth> cannot be a fraction number (带小数位), but %s !\n' % bottle * growth
first_outdepth = indepth + growth if self.first else growth
self.bn1_b = nn.BatchNorm2d(indepth)
self.conv1_b = nn.Conv2d(indepth, interdepth, 1, stride=1, padding=0, dilation=1, bias=False)
self.bn1 = nn.BatchNorm2d(interdepth)
self.conv1 = nn.Conv2d(interdepth, first_outdepth, 3, stride=2, padding=1, bias=False, dilation=1)
self.bn2_b = nn.BatchNorm2d(indepth + growth)
self.conv2_b = nn.Conv2d(indepth + growth, interdepth, 1, stride=1, padding=0, dilation=1, bias=False)
self.bn2 = nn.BatchNorm2d(interdepth)
self.conv2 = nn.Conv2d(interdepth, growth, 3, stride=2, padding=1, bias=False, dilation=1)
self.bn3_b = nn.BatchNorm2d(indepth + growth)
self.conv3_b = nn.Conv2d(indepth + growth, interdepth, 1, stride=1, padding=0, dilation=1, bias=False)
self.bn3 = nn.BatchNorm2d(interdepth)
self.deconv3 = nn.ConvTranspose2d(interdepth, growth, 3, stride=2, padding=1,
output_padding=1, bias=False, dilation=1)
self.bn4_b = nn.BatchNorm2d(indepth + growth)
self.conv4_b = nn.Conv2d(indepth + growth, interdepth, 1, stride=1, padding=0, dilation=1, bias=False)
self.bn4 = nn.BatchNorm2d(interdepth)
self.deconv4 = nn.ConvTranspose2d(interdepth, growth, 3, stride=2, padding=1,
output_padding=1, bias=False, dilation=1)
elif bottle == 0:
first_outdepth = indepth + growth if self.first else growth
self.bn1 = nn.BatchNorm2d(indepth)
self.conv1 = nn.Conv2d(indepth, first_outdepth, 3, stride=2, padding=1, bias=False, dilation=1)
self.bn2 = nn.BatchNorm2d(indepth + growth)
self.conv2 = nn.Conv2d(indepth + growth, growth, 3, stride=2, padding=1, bias=False, dilation=1)
self.bn3 = nn.BatchNorm2d(indepth + growth)
self.deconv3 = nn.ConvTranspose2d(indepth + growth, growth, 3, stride=2, padding=1,
output_padding=1, bias=False, dilation=1)
self.bn4 = nn.BatchNorm2d(indepth + growth)
self.deconv4 = nn.ConvTranspose2d(indepth + growth, growth, 3, stride=2, padding=1,
output_padding=1, bias=False, dilation=1)
else:
raise NotImplementedError('<bottle> should be >= 0, but %s' % bottle)
if self.classify > 0:
self.classifier = nn.Sequential(
nn.BatchNorm2d(indepth + growth),
Activate(active),
AdaAvgPool(),
ViewLayer(dim=-1),
nn.Linear(indepth + growth, nclass))
if self.after:
if self.down and self.reduce != 1:
outdepth = int(math.floor((indepth + growth) * reduce))
self.down_res4 = self.trans_func(indepth + growth, outdepth)
self.down_res3 = self.trans_func(indepth + growth, outdepth)
self.down_res2 = self.trans_func(indepth + growth, outdepth)
self.down_res1 = self.trans_func(indepth + growth, outdepth)
else:
if self.last_down:
outdepth = indepth + growth + last_expand
if self.last_branch >= 1:
self.down_last4 = self.trans_func(indepth + growth, outdepth)
if self.last_branch >= 2:
self.down_last3 = self.trans_func(indepth + growth, outdepth)
if self.last_branch >= 3:
self.down_last2 = self.trans_func(indepth + growth, outdepth)
if self.last_branch >= 4:
self.down_last1 = self.trans_func(indepth + growth, outdepth)
else:
if self.classify > 0 and self.last_branch == 4:
# 最后一个Couple的中间层被当做branch输出而对接在Summary上.
# 因此，删除此Couple自带的Classifier,以免与Summary中的Classifier重复.
delattr(self, 'classifier')
self.classify = 0
print('\nNote: 1 xfc will be deleted because of duplicated with the lfc!\n')
def __init__(self,
branch=3,
rock='A',
depth=16,
blockexp=(1, 1),
stages=3,
blocks=('D', 'D', 'S'),
slink=('A', 'A', 'A'),
expand=(1, 2, 4),
layers=(2, 2, 2),
dfunc=('C', 'C', 'C'),
classify=(1, 1, 1),
last_down=True,
last_expand=1,
nclass=10):
super(CifarAEResNetMix, self).__init__()
assert stages <= min(len(blocks), len(slink), len(expand),
len(layers), len(dfunc), len(classify)), \
'Hyper Pameters Not Enough to Match Stages:%s!' % stages
assert sorted(blocks[:stages]) == list(blocks[:stages]), \
'DoubleCouple must be ahead of SingleCouple! %s' % blocks[:stages]
assert stages <= 4, 'cifar stages should < 4'
self.branch = branch
self.rock = self.rocker[rock]
self.depth = depth
self.expand = expand
self.layers = layers
self.dfunc = dfunc
self.classify = classify
self.last_down = last_down
self.last_expand = last_expand
self.stages = stages
self.blocks = blocks
self.slink = slink
self.nclass = nclass
self.after = [True for _ in range(stages - 1)] + [False]
fc_indepth = depth * expand[stages - 1]
if self.last_down:
fc_indepth *= last_expand
self.layer0 = self.rock(depth=depth, branch=branch, dataset='cifar')
if stages >= 1:
self.stage1 = self._make_aelayer(self.couple[blocks[0]],
layers[0],
expand[0] * depth,
slink[0],
self.after[0],
last_down,
last_expand,
dfunc=dfunc[0],
cfy=classify[0],
blockexp=blockexp)
if stages >= 2:
self.stage2 = self._make_aelayer(self.couple[blocks[1]],
layers[1],
expand[1] * depth,
slink[1],
self.after[1],
last_down,
last_expand,
dfunc=dfunc[1],
cfy=classify[1],
blockexp=blockexp)
if stages >= 3:
self.stage3 = self._make_aelayer(self.couple[blocks[2]],
layers[2],
expand[2] * depth,
slink[2],
self.after[2],
last_down,
last_expand,
dfunc=dfunc[2],
cfy=classify[2],
blockexp=blockexp)
if stages >= 4:
self.stage4 = self._make_aelayer(self.couple[blocks[3]],
layers[3],
expand[3] * depth,
slink[3],
self.after[3],
last_down,
last_expand,
dfunc=dfunc[3],
cfy=classify[3],
blockexp=blockexp)
self.classifier = nn.Sequential(nn.BatchNorm2d(fc_indepth), nn.ReLU(),
AdaAvgPool(), ViewLayer(dim=-1),
nn.Linear(fc_indepth, nclass))
self._init_params()
def __init__(self,
block='P',
nblocks=(3, 3, 3),
inplanes=16,
bottle=1,
active='fx',
dataset='cifar10'):
""" Constructor
Args:
bottle: =1 a ResPlain block; >1 a dilate ResBottle block; <1, a shrink ResBottle block.
"""
super(XResNet, self).__init__()
assert block in self.residual.keys(
), 'Unknown residual block: %s.' % block
assert dataset in self.datasets.keys(
), 'Unsupported dataset: %s.' % dataset
assert len(nblocks) == [
4, 3
][dataset != 'imagenet'], 'Assure <nblocks> match with dataset.'
block = self.residual[block]
self.block = block
self.nblocks = nblocks
self.inplanes = inplanes
self.bottle = bottle
self.active = active
self.dataset = dataset
nlabels = self.datasets[dataset]
nplanes = [inplanes, 2 * inplanes, 4 * inplanes, 8 * inplanes]
nblocks = nblocks if len(
nblocks) >= 4 else list(nblocks) + [-1] # 添加伪数，防止越界
nbottls = [bottle] * len(nblocks)
self.preproc = PreProc(3, nplanes[0], dataset)
self.stage_1 = self._make_layer(block,
nplanes[0],
nblocks[0],
nbottls[0],
stride=1)
self.stage_2 = self._make_layer(block,
nplanes[1],
nblocks[1],
nbottls[1],
stride=2)
self.stage_3 = self._make_layer(block,
nplanes[2],
nblocks[2],
nbottls[2],
stride=2)
self.stage_4 = [
self._make_layer(block,
nplanes[3],
nblocks[3],
nbottls[3],
stride=2),
ReturnX()
][dataset != 'imagenet']
out_planes = [nplanes[-1], nplanes[-2]][dataset != 'imagenet']
self.squeeze = nn.Sequential(
nn.BatchNorm2d(out_planes * block.expansion),
nn.ReLU(inplace=True), AdaAvgPool(), ViewLayer())
self.classifier = nn.Linear(out_planes * block.expansion, nlabels)
self._init_params()
def __init__(self,
indepth,
insize,
branch,
active='relu',
pool='avg',
nclass=1000,
method='split'):
super(RockSummary, self).__init__()
assert len(
indepth) == branch, '各输出分支的通道数必须全部给定，len of <indepth> == branch.'
assert method in self.METHOD, 'Unknown <methond> %s.' % method
if method in ['convf', 'convs', 'convt']:
assert insize >= 1, '进行卷积分类，输入特征图尺寸<insize> >= 1.'
if method == 'convf':
assert branch >= 1, '对last-1进行卷积分类，输出分支<last-branch>>=1'
elif method == 'convs':
assert branch >= 2, '对last-2进行卷积分类，输出分支<last-branch>必须>=2'
elif method == 'convt':
assert branch >= 3, '对last-3进行卷积分类，输出分支<last-branch>必须>=3'
self.indepth = indepth
self.branch = branch
self.active_fc = True
self.nclass = nclass
insize = int(insize)
self.insize = insize
self.method = method
if method == 'split':
for b in range(1, branch + 1):
layer = nn.Sequential(nn.BatchNorm2d(indepth[b - 1]),
Activate(active), AdaAvgPool(),
ViewLayer(),
nn.Linear(indepth[b - 1], nclass))
setattr(self, 'classifier%s' % b, layer)
elif method == 'merge':
for b in range(1, branch + 1):
layer = nn.Sequential(nn.BatchNorm2d(indepth[b - 1]),
Activate(active), AdaAvgPool(),
ViewLayer())
setattr(self, 'pool_view%s' % b, layer)
self.classifier = nn.Linear(sum(indepth), nclass)
elif method == 'convx':
for b in range(1, branch + 1):
ksize = int(insize * ((1 / 2)**(b - 1)))
layer = ConvClassifier(indepth[b - 1],
nclass,
ksize=ksize,
stride=1,
padding=0)
setattr(self, 'classifier%s' % b, layer)
elif method == 'convf':
self.classifier = ConvClassifier(indepth[0],
nclass,
ksize=insize // 1,
stride=1,
padding=0)
elif method == 'convs':
self.classifier = ConvClassifier(indepth[1],
nclass,
ksize=insize // 2,
stride=1,
padding=0)
elif method == 'convt':
self.classifier = ConvClassifier(indepth[2],
nclass,
ksize=insize // 4,
stride=1,
padding=0)
else:
raise NotImplementedError('Unknow <method> for SummaryBlock, %s' %
method)
