def sample_latent(self, batch_size):
syntax_latent = to_var(torch.randn([batch_size, self.latent_size]))
semantic_latent = to_var(torch.randn([batch_size, self.latent_size]))
return {
"syn_z": syntax_latent,
"sem_z": semantic_latent,
}
def sampling(mean, logv):
if is_sampling:
std = torch.exp(0.5 * logv)
z = to_var(torch.randn([batch_size, self.latent_size]))
z = z * std + mean
else:
z = mean
return z
def hidden_to_latent(self, ret, is_sampling):
hidden = ret['hidden']
batch_size = hidden.size(1)
hidden = hidden.permute(1, 0, 2).contiguous()
if self.hidden_factor > 1:
hidden = hidden.view(batch_size,
self.args.enc_hidden_dim * self.hidden_factor)
else:
hidden = hidden.squeeze()
mean = self.hidden2mean(hidden)
logv = self.hidden2logv(hidden)
if is_sampling:
std = torch.exp(0.5 * logv)
z = to_var(torch.randn([batch_size, self.latent_size]))
z = z * std + mean
else:
z = mean
ret["latent"] = z
ret["mean"] = mean
ret['logv'] = logv
return ret
def sample_latent(self, batch_size):
z = to_var(torch.randn([batch_size, self.latent_size]))
return {"latent": z}