def test_tagconv():
g = dgl.DGLGraph(nx.path_graph(3))
ctx = F.ctx()
adj = g.adjacency_matrix(ctx=ctx)
norm = th.pow(g.in_degrees().float(), -0.5)
conv = nn.TAGConv(5, 2, bias=True)
conv = conv.to(ctx)
print(conv)
# test#1: basic
h0 = F.ones((3, 5))
h1 = conv(g, h0)
assert len(g.ndata) == 0
assert len(g.edata) == 0
shp = norm.shape + (1, ) * (h0.dim() - 1)
norm = th.reshape(norm, shp).to(ctx)
assert F.allclose(h1, _S2AXWb(adj, norm, h0, conv.lin.weight,
conv.lin.bias))
conv = nn.TAGConv(5, 2)
conv = conv.to(ctx)
# test#2: basic
h0 = F.ones((3, 5))
h1 = conv(g, h0)
assert h1.shape[-1] == 2
# test reset_parameters
old_weight = deepcopy(conv.lin.weight.data)
conv.reset_parameters()
new_weight = conv.lin.weight.data
assert not F.allclose(old_weight, new_weight)
def test_tagconv_e_weight(g, idtype):
ctx = F.ctx()
g = g.astype(idtype).to(ctx)
conv = nn.TAGConv(5, 5, bias=True)
conv = conv.to(ctx)
feat = F.randn((g.number_of_nodes(), 5))
eweight = F.ones((g.num_edges(), ))
conv = conv.to(ctx)
h = conv(g, feat, edge_weight=eweight)
assert h.shape[-1] == 5
def __init__(self, in_dim, hidden_dim, n_classes, hidden_layers, ctype,
hops, readout, activation_func, dropout, grid, device):
super(Classifier, self).__init__()
self.device = device
self.readout = readout
self.layers = nn.ModuleList()
self.batch_norms = nn.ModuleList()
self.grid = grid
# input layer
if ctype == 'tagconv':
self.layers.append(
conv.TAGConv(in_dim,
hidden_dim,
hops,
activation=activation_func))
else:
self.layers.append(
conv.SGConv(in_dim,
hidden_dim,
hops,
cached=False,
norm=activation_func))
self.batch_norms.append(nn.BatchNorm1d(hidden_dim))
# hidden layers
for k in range(0, hidden_layers):
if ctype == 'tagconv':
self.layers.append(
conv.TAGConv(hidden_dim,
hidden_dim,
hops,
activation=activation_func))
else:
self.layers.append(
conv.SGConv(hidden_dim,
hidden_dim,
hops,
cached=False,
norm=activation_func))
self.batch_norms.append(nn.BatchNorm1d(hidden_dim))
# dropout layer
self.dropout = nn.Dropout(p=dropout)
# last layer
if self.readout == 'max':
self.readout_fcn = conv.MaxPooling()
elif self.readout == 'mean':
self.readout_fcn = conv.AvgPooling()
elif self.readout == 'sum':
self.readout_fcn = conv.SumPooling()
elif self.readout == 'gap':
self.readout_fcn = conv.GlobalAttentionPooling(
nn.Linear(hidden_dim, 1), nn.Linear(hidden_dim,
hidden_dim * 2))
else:
self.readout_fcn = SppPooling(hidden_dim, self.grid)
if self.readout == 'spp':
self.classify = nn.Sequential(
nn.Dropout(),
nn.Linear(hidden_dim * self.grid * self.grid, hidden_dim * 2),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(2 * hidden_dim, 2 * hidden_dim),
nn.ReLU(inplace=True),
nn.Linear(2 * hidden_dim, n_classes),
)
else:
var = hidden_dim
if self.readout == 'gap':
var *= 2
self.classify = nn.Linear(var, n_classes)
def __init__(self, in_dim, hidden_dim, embed_dim, hidden_layers, hops,
readout, activation_func, dropout, local, norm, grid, K,
device):
super(Classifier, self).__init__()
self.device = device
self.readout = readout
self.layers = nn.ModuleList()
self.batch_norms = nn.ModuleList()
self.grid = grid
self.K = K
self.hidden_dim = hidden_dim
self.local = local
self.norm = norm
self.layers.append(
conv.TAGConv(in_dim, hidden_dim, hops, activation=activation_func))
# hidden layers
for k in range(0, hidden_layers):
self.layers.append(
conv.TAGConv(hidden_dim,
hidden_dim,
hops,
activation=activation_func))
# dropout layer
self.dropout = nn.Dropout(p=dropout)
if self.local:
return
# readout layer
if self.readout == 'max':
self.readout_fcn = conv.MaxPooling()
elif self.readout == 'mean':
self.readout_fcn = conv.AvgPooling()
elif self.readout == 'sum':
self.readout_fcn = conv.SumPooling()
elif self.readout == 'gap':
self.readout_fcn = conv.GlobalAttentionPooling(
nn.Linear(hidden_dim, 1), nn.Linear(hidden_dim,
hidden_dim * 2))
elif self.readout == 'sort':
self.readout_fcn = conv.SortPooling(self.K)
elif self.readout == 'set':
self.readout_fcn = conv.Set2Set(hidden_dim, 2, 1)
elif self.readout == 'cov':
self.readout_fcn = CovPooling(hidden_dim)
else:
self.readout_fcn = SppPooling(hidden_dim, self.grid)
if self.readout == 'spp':
self.embed = nn.Sequential(
nn.Dropout(),
nn.Linear(hidden_dim * self.grid * self.grid, hidden_dim * 2),
nn.ReLU(inplace=True), nn.Dropout(),
nn.Linear(2 * hidden_dim, 2 * hidden_dim),
nn.ReLU(inplace=True), nn.Linear(2 * hidden_dim, embed_dim))
elif self.readout == 'sort':
self.embed = nn.Sequential(
#nn.Dropout(),
nn.Linear(hidden_dim * self.K, embed_dim))
elif self.readout == 'cov':
self.embed = nn.Sequential(
nn.Dropout(),
nn.Linear(int(((hidden_dim + 1) * hidden_dim) / 2), embed_dim))
else:
var = hidden_dim
if self.readout == 'gap' or self.readout == 'set':
var *= 2
self.embed = nn.Linear(var, embed_dim)
def __init__(self, in_dim, hidden_dim, n_classes, hidden_layers, ctype,
hops, readout, activation_func, dropout, grid, K, norm,
device):
super(Classifier, self).__init__()
self.device = device
self.readout = readout
self.layers = nn.ModuleList()
self.n_layers = nn.ModuleList()
self.grid = grid
self.K = K
self.hidden_dim = hidden_dim
self.norm = norm
self.mish = Mish()
# input layer
if ctype == 'tagconv':
self.layers.append(
conv.TAGConv(in_dim,
hidden_dim,
hops,
activation=activation_func))
else:
self.layers.append(
conv.SGConv(in_dim,
hidden_dim,
hops,
cached=False,
norm=activation_func))
if self.norm == 'batch':
self.n_layers.append(nn.BatchNorm1d(hidden_dim))
elif self.norm == 'layer':
self.n_layers.append(
nn.LayerNorm(hidden_dim, elementwise_affine=False))
elif self.norm == 'group':
self.n_layers.append(nn.GroupNorm(16, hidden_dim))
elif self.norm == 'instance':
self.n_layers.append(nn.InstanceNorm1d(hidden_dim))
else:
self.n_layers.append(GraphNorm(hidden_dim, affine=False))
# hidden layers
for k in range(0, hidden_layers):
if ctype == 'tagconv':
self.layers.append(
conv.TAGConv(hidden_dim,
hidden_dim,
hops,
activation=activation_func))
else:
self.layers.append(
conv.SGConv(hidden_dim,
hidden_dim,
hops,
cached=False,
norm=activation_func))
if self.norm == 'batch':
self.n_layers.append(nn.BatchNorm1d(hidden_dim))
elif self.norm == 'layer':
self.n_layers.append(
nn.LayerNorm(hidden_dim, elementwise_affine=False))
elif self.norm == 'group':
self.n_layers.append(nn.GroupNorm(16, hidden_dim))
elif self.norm == 'instance':
self.n_layers.append(nn.InstanceNorm1d(hidden_dim))
else:
self.n_layers.append(GraphNorm(hidden_dim, affine=False))
# dropout layer
self.dropout = nn.Dropout(p=dropout)
# last layer
if self.readout == 'max':
self.readout_fcn = conv.MaxPooling()
elif self.readout == 'mean':
self.readout_fcn = conv.AvgPooling()
elif self.readout == 'sum':
self.readout_fcn = conv.SumPooling()
elif self.readout == 'gap':
self.readout_fcn = conv.GlobalAttentionPooling(
nn.Linear(hidden_dim, 1), nn.Linear(hidden_dim,
hidden_dim * 2))
elif self.readout == 'sort':
self.readout_fcn = conv.SortPooling(self.K)
elif self.readout == 'set':
self.readout_fcn = conv.Set2Set(hidden_dim, 2, 1)
elif self.readout == 'cov':
self.readout_fcn = CovPooling(hidden_dim)
else:
self.readout_fcn = SppPooling(hidden_dim, self.grid)
if self.readout == 'spp':
self.classify = nn.Sequential(
nn.Dropout(),
nn.Linear(hidden_dim * self.grid * self.grid, hidden_dim * 2),
nn.ReLU(inplace=True), nn.Dropout(),
nn.Linear(2 * hidden_dim, 2 * hidden_dim),
nn.ReLU(inplace=True), nn.Linear(2 * hidden_dim, n_classes))
elif self.readout == 'sort':
self.classify = nn.Sequential(
#nn.Dropout(),
nn.Linear(hidden_dim * self.K, n_classes))
elif self.readout == 'cov':
self.classify = nn.Sequential(
nn.Dropout(),
nn.Linear(int(((hidden_dim + 1) * hidden_dim) / 2), n_classes))
else:
var = hidden_dim
if self.readout == 'gap' or self.readout == 'set':
var *= 2
self.classify = nn.Linear(var, n_classes)
