def trainD(self, idf_label, y, idf_unlabel):
x_label, x_unlabel, y = self.make_input(*idf_label), self.make_input(
*idf_unlabel), Variable(y, requires_grad=False)
if self.args.cuda:
x_label, x_unlabel, y = x_label.cuda(), x_unlabel.cuda(), y.cuda()
output_label, (mom_un, output_unlabel), output_fake = self.D(
x_label, cuda=self.args.cuda), self.D(
x_unlabel, cuda=self.args.cuda,
feature=True), self.D(self.G(x_unlabel.size()[0],
cuda=self.args.cuda).view(
x_unlabel.size()).detach(),
cuda=self.args.cuda)
logz_label, logz_unlabel, logz_fake = log_sum_exp(
output_label), log_sum_exp(output_unlabel), log_sum_exp(
output_fake) # log ∑e^x_i
prob_label = torch.gather(output_label, 1,
y.unsqueeze(1)) # log e^x_label = x_label
loss_supervised = -torch.mean(prob_label) + torch.mean(logz_label)
loss_unsupervised = 0.5 * (
-torch.mean(logz_unlabel) +
torch.mean(F.softplus(logz_unlabel)) + # real_data: log Z/(1+Z)
torch.mean(F.softplus(logz_fake))) # fake_data: log 1/(1+Z)
entropy = -torch.mean(
F.softmax(output_unlabel, dim=1) *
F.log_softmax(output_unlabel, dim=1))
pt = pull_away_term(mom_un)
loss = loss_supervised + self.args.unlabel_weight * loss_unsupervised + entropy + pt
acc = torch.mean((output_label.max(1)[1] == y).float())
self.Doptim.zero_grad()
loss.backward()
self.Doptim.step()
return loss_supervised.data.cpu().numpy(), loss_unsupervised.data.cpu(
).numpy(), acc
def trainD(self, idf_label, y, idf_unlabel):
x_label, x_unlabel, y = self.make_input(*idf_label), self.make_input(
*idf_unlabel), y
output_label, (mom_un, output_unlabel), output_fake = self.D(
x_label, cuda=self.args.cuda), self.D(
x_unlabel, cuda=self.args.cuda, feature=True), self.D(
tf.reshape(self.G(x_unlabel.shape[0], cuda=self.args.cuda),
x_unlabel.shape),
cuda=self.args.cuda)
logz_label, logz_unlabel, logz_fake = log_sum_exp(
output_label), log_sum_exp(output_unlabel), log_sum_exp(
output_fake) # log ∑e^x_i
prob_label = tf.gather(output_label, 1,
tf.expand_dims(y, 1)) # log e^x_label = x_label
loss_supervised = -tf.reduce_mean(prob_label) + tf.reduce_mean(
logz_label)
loss_unsupervised = 0.5 * (
-tf.reduce_mean(logz_unlabel) + tf.reduce_mean(
tf.math.softplus(logz_unlabel)) + # real_data: log Z/(1+Z)
tf.reduce_mean(tf.math.softplus(logz_fake))
) # fake_data: log 1/(1+Z)
entropy = -tf.reduce_mean(
tf.nn.softmax(output_unlabel, axis=1) *
tf.nn.log_softmax(output_unlabel, axis=1))
pt = pull_away_term(mom_un)
loss = loss_supervised + self.args.unlabel_weight * loss_unsupervised + entropy + pt
ac = tf.reduce_mean(
tf.cast(tf.argmax(input=output_label, axis=1) == y,
dtype=tf.float32))
return loss_supervised.numpy(), loss_unsupervised.numpy(), loss, ac
def trainG(self, idf_unlabel):
x_unlabel = self.make_input(*idf_unlabel)
fake = tf.reshape(self.G(x_unlabel.shape[0], cuda=self.args.cuda),
x_unlabel.shape)
mom_gen, output_fake = self.D(fake, feature=True, cuda=self.args.cuda)
mom_unlabel, output_unlabel = self.D(x_unlabel,
feature=True,
cuda=self.args.cuda)
loss_pt = pull_away_term(mom_gen)
mom_gen = tf.raw_ops.Mean(input=mom_gen, axis=0)
mom_unlabel = tf.raw_ops.Mean(input=mom_unlabel, axis=0)
loss_fm = tf.reduce_mean(tf.abs(mom_gen - mom_unlabel))
loss = loss_fm + loss_pt
return loss
def trainG(self, idf_unlabel):
x_unlabel = self.make_input(*idf_unlabel)
if self.args.cuda:
x_unlabel = x_unlabel.cuda()
fake = self.G(x_unlabel.size()[0], cuda = self.args.cuda).view(x_unlabel.size())
mom_gen, output_fake = self.D(fake, feature=True, cuda=self.args.cuda)
mom_unlabel, output_unlabel = self.D(x_unlabel, feature=True, cuda=self.args.cuda)
loss_pt = pull_away_term(mom_gen)
mom_gen = torch.mean(mom_gen, dim = 0)
mom_unlabel = torch.mean(mom_unlabel, dim = 0)
loss_fm = torch.mean(torch.abs(mom_gen - mom_unlabel))
loss = loss_fm + loss_pt
self.Goptim.zero_grad()
self.Doptim.zero_grad()
loss.backward()
self.Goptim.step()
return loss.data.cpu().numpy()