def __init__(self, opt):
super(DeepGCN, self).__init__()
channels = opt.n_filters
act = opt.act
norm = opt.norm
bias = opt.bias
conv = opt.conv
heads = opt.n_heads
c_growth = 0
self.n_blocks = opt.n_blocks
self.head = GraphConv(opt.in_channels, channels, conv, act, norm, bias, heads)
res_scale = 1 if opt.block.lower() == 'res' else 0
if opt.block.lower() == 'dense':
c_growth = channels
self.backbone = MultiSeq(*[DenseGraphBlock(channels+i*c_growth, c_growth, conv, act, norm, bias, heads)
for i in range(self.n_blocks-1)])
else:
self.backbone = MultiSeq(*[ResGraphBlock(channels, conv, act, norm, bias, heads, res_scale)
for _ in range(self.n_blocks-1)])
fusion_dims = int(channels * self.n_blocks + c_growth * ((1 + self.n_blocks - 1) * (self.n_blocks - 1) / 2))
self.fusion_block = MLP([fusion_dims, 1024], act, None, bias)
self.prediction = Seq(*[MLP([1+fusion_dims, 512], act, norm, bias), torch.nn.Dropout(p=opt.dropout),
MLP([512, 256], act, norm, bias), torch.nn.Dropout(p=opt.dropout),
MLP([256, opt.n_classes], None, None, bias)])
self.model_init()
def __init__(self, opt):
super(DynDeepGCN, self).__init__()
channels = opt.n_filters
k = opt.kernel_size
act = opt.act
norm = opt.norm
knn = opt.knn
bias = opt.bias
epsilon = opt.epsilon
# TODO: stochastic
stochastic = False
conv = opt.conv
c_growth = channels
self.n_blocks = opt.n_blocks
self.head = GraphConv(opt.in_channels, channels, conv, act, norm, bias)
if opt.block.lower() == 'res':
self.backbone = MultiSeq(*[
ResDynBlock(channels,
k,
i % 8 + 1,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon,
knn=knn) for i in range(self.n_blocks - 1)
])
elif opt.block.lower() == 'dense':
self.backbone = MultiSeq(*[
DenseDynBlock(channels,
k,
i % 8 + 1,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon,
knn=knn) for i in range(self.n_blocks - 1)
])
else:
raise NotImplementedError(
'{} is not implemented. Please check.\n'.format(opt.block))
self.fusion_block = MLP(
[channels + c_growth * (self.n_blocks - 1), 1024], act, None, bias)
self.prediction = MultiSeq(*[
MLP([1 + channels + c_growth *
(self.n_blocks - 1), 512, 256], act, None, bias),
MLP([256, opt.n_classes], None, None, bias)
])
self.model_init()
def __init__(self, opt):
super(SparseDeepGCN, self).__init__()
channels = opt.n_filters
k = opt.kernel_size
act = opt.act
norm = opt.norm
bias = opt.bias
epsilon = opt.epsilon
stochastic = opt.stochastic
conv = opt.conv
c_growth = channels
self.n_blocks = opt.n_blocks
self.knn = DilatedKnnGraph(k, 1, stochastic, epsilon)
self.head = GraphConv(opt.in_channels, channels, conv, act, norm, bias)
if opt.block.lower() == 'res':
self.backbone = MultiSeq(*[
ResDynBlock(channels,
k,
1 + i,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon) for i in range(self.n_blocks - 1)
])
elif opt.block.lower() == 'dense':
self.backbone = MultiSeq(*[
DenseDynBlock(channels,
k,
1 + i,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon)
for i in range(self.n_blocks - 1)
])
else:
raise NotImplementedError(
'{} is not implemented. Please check.\n'.format(opt.block))
self.fusion_block = MLP(
[channels + c_growth * (self.n_blocks - 1), 1024], act, norm, bias)
self.prediction = MultiSeq(*[
MLP([channels + c_growth *
(self.n_blocks - 1) + 1024, 512], act, norm, bias),
MLP([512, 256], act, norm, bias),
torch.nn.Dropout(p=opt.dropout),
MLP([256, opt.n_classes], None, None, bias)
])
self.model_init()
def __init__(self, opt):
super(SparseDeepGCN, self).__init__()
channels = opt.n_filters
k = opt.k
act = opt.act
norm = opt.norm
bias = opt.bias
epsilon = opt.epsilon
stochastic = opt.stochastic
conv = opt.conv
c_growth = channels
self.n_blocks = opt.n_blocks
self.knn = DilatedKnnGraph(k, 1, stochastic, epsilon)
self.head = GraphConv(opt.in_channels, channels, conv, act, norm, bias)
if opt.block.lower() == 'res':
self.backbone = MultiSeq(*[
ResDynBlock(channels,
k,
1 + i,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon) for i in range(self.n_blocks - 1)
])
fusion_dims = int(channels + c_growth * (self.n_blocks - 1))
elif opt.block.lower() == 'dense':
self.backbone = MultiSeq(*[
DenseDynBlock(channels + c_growth * i,
c_growth,
k,
1 + i,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon)
for i in range(self.n_blocks - 1)
])
fusion_dims = int((channels + channels + c_growth *
(self.n_blocks - 1)) * self.n_blocks // 2)
else:
# Use PlainGCN without skip connection and dilated convolution.
stochastic = False
self.backbone = MultiSeq(*[
PlainDynBlock(channels,
k,
1,
conv,
act,
norm,
bias,
stochastic=stochastic,
epsilon=epsilon)
for i in range(self.n_blocks - 1)
])
fusion_dims = int(channels + c_growth * (self.n_blocks - 1))
self.fusion_block = MLP([fusion_dims, 1024], act, norm, bias)
self.prediction = MultiSeq(*[
MLP([fusion_dims + 1024, 512], act, norm, bias),
MLP([512, 256], act, norm, bias, drop=opt.dropout),
MLP([256, opt.n_classes], None, None, bias)
])
self.model_init()