else:
model.load_state_dict(torch.load('models/{}.bin'
.format(args.model
+ utils.getModelName(args)),
map_location=lambda storage, loc: storage)
)
# Samples latent and conditional codes randomly from prior
z = model.sample_z_prior(1)
c = model.sample_c_prior(1)
# Generate positive sample given z
c[0, 0], c[0, 1] = 1, 0
_, c_idx = torch.max(c, dim=1)
sample_idxs = model.sample_sentence(z, c, temp=0.1)
print('\nSentiment: {}'.format(dataset.idx2label(int(c_idx))))
print('Generated: {}'.format(dataset.idxs2sentence(sample_idxs)))
# Generate negative sample from the same z
c[0, 0], c[0, 1] = 0, 1
_, c_idx = torch.max(c, dim=1)
sample_idxs = model.sample_sentence(z, c, temp=0.8)
print('\nSentiment: {}'.format(dataset.idx2label(int(c_idx))))
print('Generated: {}'.format(dataset.idxs2sentence(sample_idxs)))
print()
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(model.vae_params, 5)
trainer.step()
trainer.zero_grad()
# anneal kl_weight
if ep > kld_start_inc and kld_weight < kld_max:
kld_weight += kld_inc
# print current state
if ep % LOG_INTERVAL == 0:
z = model.sample_z_prior(1)
c = model.sample_c_prior(1)
_, c_idx = torch.max(c, dim=1)
sample_idxs = model.sample_sentence(z, c)
sample_sent = dataset.idxs2sentence(sample_idxs)
print(
'epoch-{}; Loss: {:.4f}; Recon: {:.4f}; KL: {:.4f}; Grad_norm: {:.4f}; Code: {}'
.format(ep, loss.data[0], recon_loss.data[0], kl_loss.data[0],
grad_norm,
'Positive' if c_idx.data[0] == 0 else 'Negative'))
print('Sample: "{}"'.format(sample_sent))
print()
# save current model
if ep % SAVE_INTERVAL == 0:
save_base_vae_iter(ep)
save_base_vae()
