def __init__(self,
outsize,
batch_norm=False,
activation=-1,
usebias=True,
norm=False):
super(Dense, self).__init__()
self.fc = L.fcLayer(outsize, usebias, norm=norm)
self.batch_norm = batch_norm
self.activation = activation
if batch_norm:
self.bn = L.batch_norm()
if activation != -1:
self.act = L.activation(activation)
def initialize(self,
size,
chn,
ratio,
input_ratio=None,
stride=1,
pad='SAME',
activation=-1,
batch_norm=False,
usebias=True):
chn_big = int(chn * ratio)
chn_small = chn - chn_big
self.avgpool = AvgPool(2, 2)
self.Convhh = ConvLayer(size,
chn_big,
stride=stride,
pad=pad,
usebias=usebias)
self.Convhl = ConvLayer(size,
chn_small,
stride=stride,
pad=pad,
usebias=usebias)
self.Convlh = ConvLayer(size,
chn_big,
stride=stride,
pad=pad,
usebias=usebias)
self.Convll = ConvLayer(size,
chn_small,
stride=stride,
pad=pad,
usebias=usebias)
# bn and act
if batch_norm:
self.bn = L.batch_norm()
if activation != -1:
self.act = L.activation(activation)
self.batch_norm = batch_norm
self.activation = activation
self.ratio = ratio
self.input_ratio = ratio if input_ratio is None else input_ratio
self.stride = stride
self.chn_big = chn_big
def __init__(self,
outsize,
adj_mtx=None,
adj_fn=None,
usebias=True,
activation=-1,
batch_norm=False):
super(GraphConvLayer, self).__init__()
self.GCL = L.graphConvLayer(outsize,
adj_mtx=adj_mtx,
adj_fn=adj_fn,
usebias=usebias)
self.batch_norm = batch_norm
self.activation = activation
if batch_norm:
self.bn = L.batch_norm()
if activation != -1:
self.act = L.activation(activation)
def __init__(self,
size,
outchn,
dilation_rate=1,
stride=1,
pad='SAME',
activation=-1,
batch_norm=False,
usebias=True):
super(DeconvLayer3D, self).__init__()
self.conv = L.deconv3D(size, outchn, stride, pad, dilation_rate,
usebias)
self.batch_norm = batch_norm
self.activation = activation
if batch_norm:
self.bn = L.batch_norm()
if activation != -1:
self.act = L.activation(activation)
def __init__(self,
size,
outchn,
dilation_rate=1,
stride=1,
pad='SAME',
activation=-1,
batch_norm=False,
usebias=True):
"""
:type size: int or list[int]
:param size: Indicate the size of convolution kernel.
:type outchn: int
:param outchn: Number of output channels
:type stride: int or list[int]
:param stride: Stride number. Can be either integer or list of integers
:type pad: String
:param pad: Padding method, must be one of 'SAME', 'VALID', 'SAME_LEFT'. 'VALID' does not use auto-padding scheme. 'SAME' uses tensorflow-style auto-padding and 'SAME_LEFT' uses pytorch-style auto-padding.
:type dilation_rate: int or list[int]
:param dilation_rate: Dilation rate. Can be either integer or list of integers. When dilation_rate is larger than 1, stride should be 1.
:type usebias: bool
:param usebias: Whether to add bias term in this layer.
:param activation: Same candidates as layers3.activate
:type batch_norm: bool
:param batch_norm: Whether to use batch normalization in this layer.
"""
super(ConvLayer1D, self).__init__()
self.conv = L.conv1D(size, outchn, stride, pad, dilation_rate, usebias)
self.batch_norm = batch_norm
self.activation = activation
if batch_norm:
self.bn = L.batch_norm()
if activation != -1:
self.act = L.activation(activation)
def __init__(self,
size,
multiplier,
dilation_rate=1,
stride=1,
pad='SAME',
activation=-1,
batch_norm=False,
usebias=True,
values=None):
"""
:type size: int or list[int]
:param size: Indicate the size of convolution kernel.
:type multiplier: int
:param multiplier: Multiplier of number of output channel. (outchannel = multiplier * inchannel)
:type stride: int or list[int]
:param stride: Stride number. Can be either integer or list of integers
:type pad: String
:param pad: Padding method, must be one of 'SAME', 'VALID', 'SAME_LEFT'. 'VALID' does not use auto-padding scheme. 'SAME' uses tensorflow-style auto-padding and 'SAME_LEFT' uses pytorch-style auto-padding.
:type dilation_rate: int or list[int]
:param dilation_rate: Dilation rate. Can be either integer or list of integers. When dilation_rate is larger than 1, stride should be 1.
:type usebias: bool
:param usebias: Whether to add bias term in this layer.
:type values: list[np.array]
:param values: If the param 'values' is set, the layer will be initialized with the list of numpy array.
:param activation: Same candidates as layers3.activate
:type batch_norm: bool
:param batch_norm: Whether to use batch normalization in this layer.
"""
super(DWConvLayer, self).__init__()
self.batch_norm = batch_norm
self.activation = activation
self.values = values
if values is None:
self.dwconv = L.dwconv2D(size, multiplier, stride, pad,
dilation_rate, usebias)
if batch_norm:
self.bn = L.batch_norm()
else:
if usebias:
idx = 2
else:
idx = 1
self.dwconv = L.dwconv2D(size,
multiplier,
stride,
pad,
dilation_rate,
usebias,
values=values[:idx])
if batch_norm:
self.bn = L.batch_norm(values=values[idx:])
if activation != -1:
self.act = L.activation(activation)