Created by: ManWingloeng

问题报错：

这个是我定义的判别器的部分代码。

def wn(name=None) :
    if name is None:
        name = get_parent_function_name()
    return fluid.WeightNormParamAttr(dim=None, name=name, 
                       initializer=fluid.initializer.ConstantInitializer(1.0))

    def D(self, x, y_, name='D', is_test=False, reuse=False):
        with fluid.unique_name.guard(name+'_'):
            print("xshape: ",x.shape)
            x = dropout(x, dropout_prob=0.2, is_test=False)
            y = reshape(y_, [-1, self.y_dim, 1, 1]) #ten classes
            x = conv_cond_concat(x, y)
            #weight norm in paddlepaddle has finished
            print("convbefore!!!!!!!:",x)
            x = conv2d(x, num_filters=32, filter_size=3, 
                            param_attr=wn(name=name+'_conv1_weight_norm_param'),
                            act='lrelu', reuse=reuse)
            x = conv_cond_concat(x, y)
            x = conv2d(x, num_filters=32, filter_size=3, stride=2, 
                        param_attr=wn(name=name+'_conv2_weight_norm_param'), 
                        act='lrelu', reuse=reuse)
            print("D,con2d_2 shape:",x.shape)
            x = dropout(x, dropout_prob=0.2)
            x = conv_cond_concat(x, y)

            x = conv2d(x, num_filters=64, filter_size=3, param_attr=wn(), 
                              act='lrelu', reuse=reuse)
            x = conv_cond_concat(x, y)
            x = conv2d(x, num_filters=64, filter_size=3, stride=2, param_attr=wn(), 
                              act='lrelu', reuse=reuse)
            x = dropout(x, dropout_prob=0.2)
            x = conv_cond_concat(x, y)

            x = conv2d(x, num_filters=128, filter_size=3, param_attr=wn(), 
                              act='lrelu', reuse=reuse)
            x = conv_cond_concat(x, y)
            x = conv2d(x, num_filters=128, filter_size=3, param_attr=wn(), 
                              act='lrelu', reuse=reuse)
            x = conv_cond_concat(x, y)
            
            x = Global_Average_Pooling(x)
            x = flatten(x)
            x = concat(x, y_)
            #IcGAN 每一层都要concat一下
            x_logit = fc(x, 1)
            out = sigmoid(x_logit)

            return out, x_logit, x

调用D模型的program：

        with fluid.program_guard(d_program):
            # declare_data(self)
            self.inputs = fluid.layers.data(shape=image_dims, name='real_images')
            self.y = fluid.layers.data(shape=[self.y_dim], name='y')
            self.z = fluid.layers.data(shape=[self.z_dim], name='z')
            self.unlabelled_inputs = fluid.layers.data(shape=image_dims, 
                                                     name='unlabelled_images')

            print(self.inputs)
            D_real, D_real_logits, _ = self.D(self.inputs, self.y, is_test=False)
            G_train = self.G(self.z, self.y, is_test=False)
            print(G_train)
            D_fake, D_fake_logits, _ = self.D(G_train, self.y, is_test=False, reuse=True)

在实现TripleGAN的时候遇到这个问题，其中已经检查过inputs和G_train的shape是一致的，因为我需要对conv2d中的权重参数进行初始化，发现可以用WeightNormParamAttr，但是第当第二次调用即：D_fake, D_fake_logits, _ = self.D(G_train, self.y, is_test=False, reuse=True)这里一直报与之前的权重w_g的shape不一致？