Question

我试图将conv1d_transpose与VALID填充一起用于一维生成对抗网络，并不断出现以下错误：

ValueError: Incompatible shapes between op input and calculated input gradient.  Forward operation: g_Generator_Output/Generator_Output.  Input index: 2. Original input shape: (?, 1, 256, 32).  Calculated input gradient shape: (?, 1, 253, 32)

我的理解是输出宽度应使用：

W_out =（W_in-1）*步幅+过滤宽度

在我的情况下为256 =（253-1）* 1 + 4

按照错误消息的建议，我尝试将W_out设置为253，但这会返回以下错误

ValueError: Incompatible shapes between op input and calculated input gradient.  Forward operation: g_gen_Deconv2/gen_Deconv2.  Input index: 2. Original input shape: (?, 1, 253, 32).  Calculated input gradient shape: (?, 1, 250, 32)

相关代码如下，其中self.dc [i-1]是反卷积的输入：

  for i in range(1,len(deconv_shape)):
        with tf.name_scope(self.name+'_'+deconv_shape[i][-1]):
            batch_size=tf.shape(self.dc[i-1])[0]

            if deconv_shape[i][3]=='VALID':
                W_out=(self.dc[i-1].shape[1].value-1)*deconv_shape[i][2]+deconv_shape[i][0][0]
                #W_out=253


            deconv_shape[i][1]=tf.stack([batch_size,W_out,deconv_shape[i][1][2]])
            deconv_shape[i][0]=tf.get_variable(self.name+'_'+deconv_shape[i][-1]+'_filters', deconv_shape[i][0], initializer=tf.random_normal_initializer(stddev=0.02))
            self.dc.append(tf.contrib.nn.conv1d_transpose(self.dc[i-1],*deconv_shape[i]))
            if BATCH_NORMALISE:
                self.dc[i]=tf.layers.batch_normalization(self.dc[i],training=self.training,name=self.name+'_'+deconv_shape[i][-1]+'_BN')                
            tf.summary.histogram('dc'+str(i),self.dc[i])

和deconv_shape [i]定义如下：

[[kernel_size,filters,infilters],[BATCH_SIZE,1,filters],strides,'VALID','NWC','gen_Deconv'+str(i)]

kernel_size = 4，filters = infilters = 32，步幅= 1

tf.contrib.nn.conv1d_transpose

0 个答案: