vae_m1.py 文件源码

def train(self, x, L=1, test=False):
        batchsize = x.data.shape[0]
        z_mean, z_ln_var = self.encoder(x, test=test, apply_f=False)
        loss = 0
        for l in xrange(L):
            # Sample z
            z = F.gaussian(z_mean, z_ln_var)

            # Compute lower bound
            log_px_z = self.log_px_z(x, z, test=test)
            log_pz = self.log_pz(z, z_mean, z_ln_var)
            log_qz_x = self.log_qz_x(z, z_mean, z_ln_var)
            lower_bound = log_px_z + log_pz - log_qz_x

            loss += -lower_bound

        loss = F.sum(loss) / L / batchsize

        self.zero_grads()
        loss.backward()
        self.update()

        if self.gpu:
            loss.to_cpu()
        return loss.data