def __init__(self, feature_dim, hidden_dim):
super(DANNDiscriminator, self).__init__()
self.ad_layer1 = nn.Linear(feature_dim, hidden_dim)
self.ad_layer2 = nn.Linear(hidden_dim, hidden_dim)
self.ad_layer3 = nn.Linear(hidden_dim, 1)
self.relu = nn.ReLU()
self.grl_layer = grl.GradientReverseLayer()
self.sigmoid = nn.Sigmoid()
self.drop_layer1 = nn.Dropout(0.5)
self.drop_layer2 = nn.Dropout(0.5)
self.ad_layer1.weight.data.normal_(0, 0.01)
self.ad_layer2.weight.data.normal_(0, 0.01)
self.ad_layer3.weight.data.normal_(0, 0.3)
self.ad_layer1.bias.data.fill_(0.0)
self.ad_layer2.bias.data.fill_(0.0)
self.ad_layer3.bias.data.fill_(0.0)
self.parameter_list = [{
"params": self.ad_layer1.parameters(),
"lr": 10
}, {
"params": self.ad_layer2.parameters(),
"lr": 10
}, {
"params": self.ad_layer3.parameters(),
"lr": 10
}]
def forward(self, inputs):
self.iter_count += 1
grl_layer = grl.GradientReverseLayer(iter_num=self.iter_count)
outputs = grl_layer(inputs)
outputs = self.drop_layer1(self.relu(self.ad_layer1(outputs)))
outputs = self.drop_layer2(self.relu(self.ad_layer2(outputs)))
outputs = self.sigmoid(self.ad_layer3(outputs))
return outputs
def forward(self, features, softmax_outputs):
self.forward_count += 1
grl_layer = grl.GradientReverseLayer(iter_num=self.forward_count)
if self.randomize:
rman_out_list = self.rmm_layer.forward([features, softmax_outputs])
rman_out = rman_out_list[0]
for rman_single_out in rman_out_list[1:]:
rman_out = torch.mul(rman_out, rman_single_out)
adv_inputs = rman_out.view(-1, rman_out.size(1))
else:
op_out = torch.bmm(softmax_outputs.unsqueeze(2),
features.unsqueeze(1))
adv_inputs = op_out.view(
-1,
softmax_outputs.size(1) * features.size(1))
adv_inputs = torch.sqrt(torch.abs(adv_inputs) + 1e-12)
adv_inputs = F.normalize(adv_inputs, p=2, dim=1)
outputs = grl_layer(adv_inputs)
outputs = self.drop_layer1(self.relu(self.ad_layer1(outputs)))
outputs = self.drop_layer2(self.relu(self.ad_layer2(outputs)))
outputs = self.sigmoid(self.ad_layer3(outputs))
return outputs
