def residual_block(m, a, b, leakiness=0.01, dimensions=2):
"""
append to a sequence module:
produce output of [identity,3x3+3x3] then add together
inputs
------
m [scn.Sequential module] network to add layers to
a [int]: number of input channels
b [int]: number of output channels
leakiness [float]: leakiness of ReLU activations
dimensions [int]: dimensions of input sparse tensor
modifies
--------
m: adds layers
"""
m.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, False)).add(
scn.Sequential().add(scn.BatchNormLeakyReLU(
a, leakiness=leakiness)).add(
scn.SubmanifoldConvolution(dimensions, a, b, 3, False)).add(
scn.BatchNormLeakyReLU(b, leakiness=leakiness)).add(
scn.SubmanifoldConvolution(
dimensions, b, b, 3, False)))).add(scn.AddTable())
def bar(nPlanes, bias):
m = scn.Sequential()
m.add(scn.BatchNormReLU(nPlanes))
m.add(scn.NetworkInNetwork(
nPlanes, nClasses,
bias)) #accumulte softmax input, only one set of biases
return m
def _make_transpose(self, transblock, planes, blocks, stride=1):
upsample = None
if stride != 1:
upsample = scn.Sequential(
scn.SparseToDense(2,self.inplanes * transblock.expansion),
nn.ConvTranspose2d(self.inplanes * transblock.expansion, planes,
kernel_size=2, stride=stride, padding=0, bias=False),
scn.DenseToSparse(2),
scn.BatchNormalization(planes)
)
elif self.inplanes * transblock.expansion != planes:
upsample = scn.Sequential(
scn.NetworkInNetwork(self.inplanes * transblock.expansion, planes, False),
scn.BatchNormalization(planes)
)
layers = []
for i in range(1, blocks):
layers.append(transblock(self.inplanes, self.inplanes * transblock.expansion))
layers.append(transblock(self.inplanes, planes, stride, upsample))
self.inplanes = planes // transblock.expansion
return scn.Sequential(*layers)
def block(self,
m,
a,
b,
dimension=3,
residual_blocks=False,
leakiness=0,
kernel_size=3,
use_batch_norm=True): # default using residual_block
if use_batch_norm:
Activation = lambda channels: scn.BatchNormLeakyReLU(
channels, leakiness=leakiness)
else:
Activation = lambda channels: scn.LeakyReLU(leakiness)
if residual_blocks: #ResNet style blocks
m.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, False)).add(
scn.Sequential().add(Activation(a)).add(
scn.SubmanifoldConvolution(dimension, a, b, kernel_size,
False)).add(Activation(b)).add(
scn.SubmanifoldConvolution(
dimension, b, b,
kernel_size,
False)))).add(
scn.AddTable())
else: #VGG style blocks
m.add(scn.Sequential().add(Activation(a)).add(
scn.SubmanifoldConvolution(dimension, a, b, kernel_size,
False)))
def residual(nIn, nOut, stride):
if stride > 1:
return scn.Convolution(dimension, nIn, nOut, 3, stride, False)
elif nIn != nOut:
return scn.NetworkInNetwork(nIn, nOut, False)
else:
return scn.Identity()
def block(m, a, b):
if residual_blocks: #ResNet style blocks
m.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, False)).add(
scn.Sequential().add(
scn.BatchNormLeakyReLU(
a,
momentum=bn_momentum,
leakiness=leakiness,
track_running_stats=track_running_stats)
).add(scn.SubmanifoldConvolution(
dimension, a, b, 3, False)).add(
scn.BatchNormLeakyReLU(
b,
momentum=bn_momentum,
leakiness=leakiness,
track_running_stats=track_running_stats)).add(
scn.SubmanifoldConvolution(
dimension, b, b, 3,
False)))).add(scn.AddTable())
else: #VGG style blocks
m.add(scn.Sequential().add(
scn.BatchNormLeakyReLU(
a,
momentum=bn_momentum,
leakiness=leakiness,
track_running_stats=track_running_stats)).add(
scn.SubmanifoldConvolution(dimension, a, b, 3,
False)))
operation = {'kernel': [1, 1, 1], 'stride': [1, 1, 1]}
return operation
def _make_skip_layer(self, inplanes, planes):
layers = scn.Sequential(
scn.NetworkInNetwork(inplanes, planes, False),
scn.BatchNormReLU(planes)
)
return layers
def block(self, n_in, n_out):
m = scn.Sequential()
if self.residual_blocks: # ResNet style blocks
m.add(scn.ConcatTable().add(
scn.Identity() if n_in ==
n_out else scn.NetworkInNetwork(n_in, n_out, False)).add(
scn.Sequential().add(
scn.BatchNormLeakyReLU(
n_in, leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, n_in, n_out, 3,
False)).add(
scn.BatchNormLeakyReLU(
n_out,
leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, n_out,
n_out, 3, False))))
m.add(scn.AddTable())
else: # VGG style blocks
m.add(scn.BatchNormLeakyReLU(n_in, leakiness=self.leakiness))
m.add(
scn.SubmanifoldConvolution(self.dimension, n_in, n_out, 3,
False))
return m
def res(m, dimension, a, b):
m.add(scn.ConcatTable()
.add(scn.Identity() if a == b else scn.NetworkInNetwork(a, b, False))
.add(scn.Sequential()
.add(scn.BatchNormReLU(a))
.add(scn.SubmanifoldConvolution(dimension, a, b, 3, False))
.add(scn.BatchNormReLU(b))
.add(scn.SubmanifoldConvolution(dimension, b, b, 3, False))))\
.add(scn.AddTable())
def block(m, a, b): # ResNet style blocks
m.add(scn.ConcatTable()
.add(scn.Identity() if a == b else scn.NetworkInNetwork(a, b, False))
.add(scn.Sequential()
.add(scn.BatchNormLeakyReLU(a, leakiness=leakiness))
.add(scn.SubmanifoldConvolution(self._dimension, a, b, 3, False))
.add(scn.BatchNormLeakyReLU(b, leakiness=leakiness))
.add(scn.SubmanifoldConvolution(self._dimension, b, b, 3, False)))
).add(scn.AddTable())
def __init__(self, inplanes, planes, stride=1, upsample=None, **kwargs):
super(TransBottleneckSparse, self).__init__()
self.conv1 = scn.NetworkInNetwork(inplanes * 4, inplanes, False)
self.bn1 = scn.BatchNormReLU(inplanes)
if upsample is not None and stride != 1:
self.conv2 = scn.Sequential(
scn.SparseToDense(2,inplanes),
nn.ConvTranspose2d(inplanes, inplanes,
kernel_size=2, stride=stride, padding=0,
output_padding=0, bias=False),
scn.DenseToSparse(2)
)
else:
self.conv2 = conv3x3_sparse(inplanes, inplanes, stride)
self.bn2 = scn.BatchNormReLU(inplanes)
self.conv3 = scn.NetworkInNetwork(inplanes, planes, False)
self.bn3 = scn.BatchNormalization(planes)
self.relu = scn.ReLU()
self.add = scn.AddTable()
self.upsample = upsample
self.stride = stride
def f(m, a, b):
m.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, self.allow_bias)).add(
scn.Sequential().add(norm_layer(
a, leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, a, b, 3, self.allow_bias)).add(
norm_layer(b, leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, b, b, 3,
self.allow_bias)))).add(scn.AddTable())
return m
def __init__(self, transblock, layers, num_classes=150):
self.inplanes = 512
super(ResNetTransposeSparse, self).__init__()
self.dense_to_sparse = scn.DenseToSparse(2)
self.add = AddSparseDense()
self.deconv1 = self._make_transpose(transblock, 256 * transblock.expansion, layers[0], stride=2)
self.deconv2 = self._make_transpose(transblock, 128 * transblock.expansion, layers[1], stride=2)
self.deconv3 = self._make_transpose(transblock, 64 * transblock.expansion, layers[2], stride=2)
self.deconv4 = self._make_transpose(transblock, 64 * transblock.expansion, layers[3], stride=2)
self.skip0 = self._make_skip_layer(128, 64 * transblock.expansion)
self.skip1 = self._make_skip_layer(256, 64 * transblock.expansion)
self.skip2 = self._make_skip_layer(512, 128 * transblock.expansion)
self.skip3 = self._make_skip_layer(1024, 256 * transblock.expansion)
self.skip4 = self._make_skip_layer(2048, 512 * transblock.expansion)
self.densify0 = scn.SparseToDense(2, 64 * transblock.expansion)
self.densify1 = scn.SparseToDense(2, 64 * transblock.expansion)
self.densify2 = scn.SparseToDense(2, 128 * transblock.expansion)
self.densify3 = scn.SparseToDense(2, 256 * transblock.expansion)
self.inplanes = 64
self.final_conv = self._make_transpose(transblock, 64 * transblock.expansion, 3)
self.final_deconv = scn.Sequential(
scn.SparseToDense(2, self.inplanes * transblock.expansion),
nn.ConvTranspose2d(self.inplanes * transblock.expansion, num_classes, kernel_size=2,
stride=2, padding=0, bias=True)
)
self.out6_conv = nn.Conv2d(2048, num_classes, kernel_size=1, stride=1, bias=True)
self.out5_conv = scn.NetworkInNetwork(256 * transblock.expansion, num_classes, True)
self.out4_conv = scn.NetworkInNetwork(128 * transblock.expansion, num_classes, True)
self.out3_conv = scn.NetworkInNetwork(64 * transblock.expansion, num_classes, True)
self.out2_conv = scn.NetworkInNetwork(64 * transblock.expansion, num_classes, True)
self.sparse_to_dense = scn.SparseToDense(2, num_classes)
def block(m, a, b):
if residual_blocks: #ResNet style blocks
m.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, False)).add(
scn.Sequential().add(scn.BatchNormReLU(a)).add(
scn.SubmanifoldConvolution(dimension, a, b, 3, False)).add(
scn.BatchNormReLU(b)).add(
scn.SubmanifoldConvolution(dimension, b, b, 3,
False)))).add(
scn.AddTable())
else: #VGG style blocks
m.add(scn.Sequential().add(scn.BatchNormReLU(a)).add(
scn.SubmanifoldConvolution(dimension, a, b, 3, False)))
def block(self, m, a, b, dimension=3, residual_blocks=False, leakiness=0): # default using residual_block
if residual_blocks: #ResNet style blocks
m.add(scn.ConcatTable()
.add(scn.Identity() if a == b else scn.NetworkInNetwork(a, b, False))
.add(scn.Sequential()
.add(scn.BatchNormLeakyReLU(a,leakiness=leakiness))
.add(scn.SubmanifoldConvolution(dimension, a, b, 3, False))
.add(scn.BatchNormLeakyReLU(b,leakiness=leakiness))
.add(scn.SubmanifoldConvolution(dimension, b, b, 3, False)))
).add(scn.AddTable())
else: #VGG style blocks
m.add(scn.Sequential()
.add(scn.BatchNormLeakyReLU(a,leakiness=leakiness))
.add(scn.SubmanifoldConvolution(dimension, a, b, 3, False)))
def _nin_block(self, module, a, b):
'''
Utility Method for attaching feature dimension reducing
BN + NetworkInNetwork blocks.
INPUTS:
- module (scn Module): network module to attach ResNet block.
- a (int): number of input feature dimension
- b (int): number of output feature dimension
RETURNS:
None (operation is in-place)
'''
module.add(scn.Sequential().add(
scn.BatchNormLeakyReLU(a, leakiness=self.leakiness)).add(
scn.NetworkInNetwork(a, b, self.allow_bias)))
def get_channel_changer_or_identity(
num_dims, sparse, input_channels, output_channels=None,
residual_changer_bias=True):
stride = np.full(num_dims, 1)
if output_channels is None or input_channels == output_channels:
return get_identity(num_dims, sparse, input_channels)
if sparse:
layer = scn.NetworkInNetwork(
input_channels, output_channels, residual_changer_bias)
else:
conv_class = get_dense_conv_class(num_dims)
layer = conv_class(
input_channels, output_channels, kernel_size=1,
bias=residual_changer_bias)
return sparse, stride, output_channels, layer
def resnet_block(dimension,
n_in,
n_out,
kernel,
leakiness=0,
computation='convolution'):
'''Build and return ResNet block
'''
assert computation in [
'submanifoldconvolution', 'convolution', 'fullconvolution',
'deconvolution'
]
if computation == 'convolution':
computation = lambda n_in, n_out: scn.Convolution(
dimension, n_in, n_out, kernel[0], kernel[1], False)
elif computation == 'submanifoldconvolution':
assert type(
kernel
) == int, f"`kernel` must be int, {type(kernel)} was provided"
computation = lambda n_in, n_out: scn.SubmanifoldConvolution(
dimension, n_in, n_out, kernel, False)
elif computation == 'deconvolution':
assert type(
kernel
) == int, f"`kernel` must be int, {type(kernel)} was provided"
computation = lambda n_in, n_out: scn.Deconvolution(
dimension, n_in, n_out, kernel, kernel, False)
else:
computation = lambda n_in, n_out: scn.FullConvolution(
dimension, n_in, n_out, kernel[0], kernel[1], False)
block = scn.Sequential()
block.add(scn.ConcatTable(
).add(scn.NetworkInNetwork(n_in, n_out, False)).add(scn.Sequential().add(
# scn.BatchNormLeakyReLU(n_in, leakiness=leakiness)
scn.LeakyReLU(leakiness)).add(computation(n_in, n_out)).add(
# scn.BatchNormLeakyReLU(n_out, leakiness=leakiness)
scn.LeakyReLU(leakiness)).add(computation(n_out, n_out)))).add(
scn.AddTable())
return block
def _resnet_block(self, module, a, b):
'''
Utility Method for attaching ResNet-Style Blocks.
INPUTS:
- module (scn Module): network module to attach ResNet block.
- a (int): number of input feature dimension
- b (int): number of output feature dimension
RETURNS:
None (operation is in-place)
'''
module.add(scn.ConcatTable().add(scn.Identity(
) if a == b else scn.NetworkInNetwork(a, b, self.allow_bias)).add(
scn.Sequential().add(
scn.BatchNormLeakyReLU(a, leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, a, b, 3, self.allow_bias)).add(
scn.BatchNormLeakyReLU(
b, leakiness=self.leakiness)).add(
scn.SubmanifoldConvolution(
self.dimension, b, b, 3,
self.allow_bias)))).add(scn.AddTable())
def bar(m, nPlanes, bias):
m.add(scn.BatchNormLeakyReLU(nPlanes, leakiness=leakiness))
m.add(scn.NetworkInNetwork(
nPlanes, nClasses,
bias)) #accumulte softmax input, only one set of biases
