def __init__(self, in_channels, out_channels,
edge_channels, activation, edge_mode,
normalize_emb,
aggr):
super(EGraphSage, self).__init__(aggr=aggr)
self.in_channels = in_channels
self.out_channels = out_channels
self.edge_channels = edge_channels
self.edge_mode = edge_mode
self.tanh = nn.Tanh()
if edge_mode == 0:
self.message_lin = nn.Linear(in_channels, out_channels)
self.attention_lin = nn.Linear(2*in_channels+edge_channels, 1)
elif edge_mode == 1:
self.message_lin = nn.Linear(in_channels+edge_channels, out_channels)
elif edge_mode == 2:
self.message_lin = nn.Linear(2*in_channels+edge_channels, out_channels)
elif edge_mode == 3:
self.message_lin = nn.Sequential(
nn.Linear(2*in_channels+edge_channels, out_channels),
get_activation(activation),
nn.Linear(out_channels, out_channels),
)
elif edge_mode == 4:
self.message_lin = nn.Linear(in_channels, out_channels*edge_channels)
elif edge_mode == 5:
self.message_lin = nn.Linear(2*in_channels, out_channels*edge_channels)
elif edge_mode == 6:
self.source_lin = nn.Linear(in_channels, 8)
self.message_lin = nn.Linear(8+edge_channels, out_channels)
elif edge_mode == 7:
self.message_lin = nn.Linear(edge_channels, out_channels)
self.agg_lin = nn.Linear(in_channels+out_channels, out_channels)
self.message_activation = get_activation(activation)
self.update_activation = get_activation(activation)
self.normalize_emb = normalize_emb
##GGNN
self.GRU = nn.GRU(out_channels, in_channels, 1)
self.Lin_GRU = nn.Linear(in_channels, out_channels)
##ResNET
self.lin_res = nn.Sequential(
nn.Linear(in_channels, out_channels),
nn.Sigmoid()
)
def build_edge_update_mlps(self, node_dim, edge_input_dim, edge_dim,
gnn_layer_num, activation):
edge_update_mlps = nn.ModuleList()
edge_update_mlp = nn.Sequential(
nn.Linear(node_dim + node_dim + edge_input_dim, edge_dim),
get_activation(activation),
)
edge_update_mlps.append(edge_update_mlp)
for l in range(1, gnn_layer_num):
edge_update_mlp = nn.Sequential(
nn.Linear(node_dim + node_dim + edge_dim, edge_dim),
get_activation(activation),
)
edge_update_mlps.append(edge_update_mlp)
return edge_update_mlps
def build_node_post_mlp(self, input_dim, output_dim, hidden_dims, dropout,
activation):
if 0 in hidden_dims:
return get_activation('none')
else:
layers = []
for hidden_dim in hidden_dims:
layer = nn.Sequential(
nn.Linear(input_dim, hidden_dim),
get_activation(activation),
nn.Dropout(dropout),
)
layers.append(layer)
input_dim = hidden_dim
layer = nn.Linear(input_dim, output_dim)
layers.append(layer)
return nn.Sequential(*layers)
def _dense_branches(x):
for idx, unit in enumerate(units):
x = Dense(units=unit, name=f'{name}_dense_{idx}')(x)
x = get_activation(activation)(x)
if use_dropout:
x = Dropout(rate=dropout_rate)(x)
x = Dense(units=output_num,
activation='softmax',
name=f'{name}_output')(x)
return x
def _deform_conv_block(x):
for idx, channel in enumerate(channels):
x = ConvOffset2D(deform_channels[idx],
name=f'{name}_deform_offset_{idx+name_offset}')(x)
x = Conv2D(filters=channel,
kernel_size=kernel_size,
strides=strides,
padding='same',
name=f'{name}_deform_conv_{idx+name_offset}')(x)
x = get_activation(activation)(x)
return x
def __init__(self, in_channels, out_channels, edge_channels, activation,
edge_mode, normalize_emb, aggr):
super(EGraphSage, self).__init__(aggr=aggr)
self.in_channels = in_channels
self.out_channels = out_channels
self.edge_channels = edge_channels
self.edge_mode = edge_mode
if edge_mode == 0:
self.message_lin = nn.Linear(in_channels, out_channels)
self.attention_lin = nn.Linear(2 * in_channels + edge_channels, 1)
elif edge_mode == 1:
self.message_lin = nn.Linear(in_channels + edge_channels,
out_channels)
elif edge_mode == 2:
self.message_lin = nn.Linear(2 * in_channels + edge_channels,
out_channels)
elif edge_mode == 3:
self.message_lin = nn.Sequential(
nn.Linear(2 * in_channels + edge_channels, out_channels),
get_activation(activation),
nn.Linear(out_channels, out_channels),
)
elif edge_mode == 4:
self.message_lin = nn.Linear(in_channels,
out_channels * edge_channels)
elif edge_mode == 5:
self.message_lin = nn.Linear(2 * in_channels,
out_channels * edge_channels)
self.agg_lin = nn.Linear(in_channels + out_channels, out_channels)
self.message_activation = get_activation(activation)
self.update_activation = get_activation(activation)
self.normalize_emb = normalize_emb
def conv2d(x,
filters,
kernel_size=3,
strides=1,
use_sn=False,
pad_type=None,
pad_size=None,
norm=None,
activation=None):
if pad_type == 'zero':
x = ZeroPadding2D(padding=pad_size)(x)
if pad_type in [None, 'zero']:
x = SN(Conv2D(filters, kernel_size=kernel_size, strides=strides))(x) if use_sn else \
Conv2D(filters, kernel_size=kernel_size, strides=strides)(x)
else:
x = SN(Conv2D(filters, kernel_size=kernel_size, strides=strides, padding=pad_type))(x) if use_sn else \
Conv2D(filters, kernel_size=kernel_size, strides=strides, padding=pad_type)(x)
# Normalization
x = get_normalization(norm)(x)
# Activation
x = get_activation(activation)(x)
return x
def define_model(self):
input_layer = Input(shape=self.config.input_shape)
x = conv2d(input_layer,
filters=64,
kernel_size=7,
strides=2,
pad_type='zero',
pad_size=3,
norm='in',
activation='lrelu')
x = ZeroPadding2D(padding=(1, 1))(x)
x = MaxPooling2D((3, 3), 2)(x)
shortcut = conv2d(x,
filters=256,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in')
for i in range(3):
x = conv2d(x,
filters=64,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in',
activation='lrelu')
x = conv2d(x,
filters=64,
kernel_size=3,
strides=1,
pad_type='same',
norm='in',
activation='lrelu')
x = conv2d(x,
filters=256,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in')
x = Add()([x, shortcut])
x = get_activation('lrelu')(x)
shortcut = x
shortcut = conv2d(x,
filters=512,
kernel_size=1,
strides=2,
pad_type='valid',
norm='in')
for i in range(4):
x = conv2d(x, filters=128, kernel_size=1, strides=2, pad_type='valid', norm='in', activation='lrelu') if i == 0 \
else conv2d(x, filters=128, kernel_size=1, strides=1, pad_type='valid', norm='in', activation='lrelu')
x = conv2d(x,
filters=128,
kernel_size=3,
strides=1,
pad_type='same',
norm='in',
activation='lrelu')
x = conv2d(x,
filters=512,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in')
x = Add()([x, shortcut])
x = get_activation('lrelu')(x)
shortcut = x
shortcut = conv2d(x,
filters=1024,
kernel_size=1,
strides=2,
pad_type='valid',
norm='in')
for i in range(6):
x = conv2d(x, filters=256, kernel_size=1, strides=2, pad_type='valid', norm='in', activation='lrelu') if i == 0 \
else conv2d(x, filters=256, kernel_size=1, strides=1, pad_type='valid', norm='in', activation='lrelu')
x = conv2d(x,
filters=256,
kernel_size=3,
strides=1,
pad_type='same',
norm='in',
activation='lrelu')
x = conv2d(x,
filters=1024,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in')
x = Add()([x, shortcut])
x = get_activation('lrelu')(x)
shortcut = x
shortcut = conv2d(x,
filters=2048,
kernel_size=1,
strides=2,
pad_type='valid',
norm='in')
for i in range(3):
x = conv2d(x, filters=512, kernel_size=1, strides=2, pad_type='valid', norm='in', activation='lrelu') if i == 0 \
else conv2d(x, filters=512, kernel_size=1, strides=1, pad_type='valid', norm='in', activation='lrelu')
x = conv2d(x,
filters=512,
kernel_size=3,
strides=1,
pad_type='same',
norm='in',
activation='lrelu')
x = conv2d(x,
filters=2048,
kernel_size=1,
strides=1,
pad_type='valid',
norm='in')
x = Add()([x, shortcut])
x = get_activation('lrelu')(x)
shortcut = x
x = GlobalAveragePooling2D()(x)
out = Dense(self.config.num_classes,
activation='softmax',
name=f'class_output')(x)
return Model(inputs=input_layer, outputs=out, name='resnet50')