poseVAE.py 文件源码

def genLossAndGradient(self):
        # establish loss
        kl_div = lasagne.layers.get_output(self.kl_loss_layer,
                                           deterministic=False)
        kl_loss = lasagne.objectives.aggregate(kl_div, mode='sum')

        # assume the reconstructed all with standard Gaussian distribution
        recons_loss = lasagne.objectives.squared_error(self.recons_var,
                                                       self.pose_input_var)
        recons_loss = recons_loss*0.5
        recons_loss = lasagne.objectives.aggregate(recons_loss, mode='sum')

        # calculate gradient
        loss = kl_loss + recons_loss
        # loss = recons_loss
        lr_var = T.fscalar('lr')
        update_params = self.encoder_params + self.decoder_params
        update_vars = lasagne.updates.adam(loss, update_params, 
                                           lr_var, self.b1)

        # compile the function
        self.train_fn = theano.function(
            [self.pose_input_var, self.noise_input_var, lr_var],
            loss,
            updates = update_vars)
        self.recons_fn = theano.function(
            [self.pose_input_var, self.noise_input_var],
            self.recons_tvar
        )
        self.encode_fn = theano.function(
            [self.pose_input_var, self.noise_input_var],
            self.z_tvar
        )

        print '[PoseVAE]function compiled'