G_losses = []
D_losses = []
mmd_list = []
series_list = np.zeros((1, seq_length))
for n in tqdm(range(num_epoch)):
# for k in range(1):
for n_batch, sample_data in enumerate(data_loader):
for d in range(D_rounds):
### TRAIN DISCRIMINATOR ON FAKE DATA
discriminator.zero_grad()
h_d = discriminator.init_hidden()
h_g = generator.init_hidden()
#Generating the noise and label data
noise_sample = Variable(noise(len(sample_data), seq_length))
#Use this line if generator outputs hidden states: dis_fake_data, (h_g_n,c_g_n) = generator.forward(noise_sample,h_g)
dis_fake_data = generator.forward(noise_sample, h_g).detach()
y_pred_fake, (h_d_n, c_d_n) = discriminator(dis_fake_data, h_d)
loss_fake = loss(y_pred_fake,
torch.zeros([len(sample_data), 1]).cuda())
loss_fake.backward()
#Train discriminator on real data
G_optimizer = optim.Adam(chain(Q.parameters(), P.parameters()),
lr=5e-5,
betas=(0.5, 0.999))
D_optimizer = optim.Adam(D.parameters(), lr=5e-5, betas=(0.5, 0.999))
criterion = nn.CrossEntropyLoss()
training_pairs = [
variables_from_pair(input_lang, output_lang, random.choice(pairs))
for i in range(int(1e5))
]
h_Q = Q.init_hidden(batch)
h_P = P.init_hidden(batch)
h_D_enc = D.init_hidden(batch)
h_D_gen = D.init_hidden(batch)
for epoch in range(int(1e5)):
x = training_pairs[epoch][0]
x = x.unsqueeze(0).cpu()
reset_grad([Q, P, D])
x = to_onehot(x.squeeze()).unsqueeze(0)
z = Variable(torch.zeros(batch, x.size(1), hidden_size))
E_x, h_E = Q(x.cpu(), h_Q.cpu())
G_z, h_G = P(z.cpu(), h_P.cpu())
