python - TensorFlow：使用相同的模型两次生成双变量

Question

我正在尝试使用自己训练的预训练模型，它看起来像：

    def Conv2D(x, channel, kernel_size, stride=1, bias=True, padding='same',name="conv2d",reuse=False):
        with tf.variable_scope(name):
            conv_out = tf.keras.layers.Conv2D(channel,kernel_size=kernel_size,strides=(stride,stride),padding=padding,use_bias=bias)(x)
            return conv_out

    class MyNet_Pretrain():
    def __init__(self):
      self.network_name = "net_pretrain"

    def _mynet_pretrain(self,l1):
      with tf.variable_scope(self.network_name) as vs:
        conv1_down = ReLU(Conv2D(l1, 32, 3, 1, True,name='conv1_down'))
        conv2_down = ReLU(Conv2D(conv1_down, 64, 3, 2, True,name='conv2_down'))
        conv3_down = ReLU(Conv2D(conv2_down, 128, 3, 2, True,name='conv3_down'))

        conv3_down_upsample = tf.image.resize_bicubic(conv3_down, [128, 128], True)
        conv1_up = ReLU(Conv2D(conv3_down_upsample, 64, 3, 1, True,name='conv1_up'))
        conv1_up_upsample = tf.image.resize_bicubic(conv1_up, [256, 256], True)
        conv2_up = ReLU(Conv2D(conv1_up_upsample, 64, 3, 1, True,name='conv2_up'))
        conv3_up = Conv2D(conv2_up, 4, 1, 1, True,name='conv3_up')

      return conv3_down, conv3_up

我想用它来计算生成图像和目标图像的特征图，如下所示：

    net_pretrain = MyNet_Pretrain()
    outputs_pretrain1 = net_pretrain._mynet_pretrain(outputs)
    outputs_pretrain2 = net_pretrain._mynet_pretrain(target_pl)

然而，当我检查所有可训练变量时，我发现这个网络的变量增加了一倍：

    [<tf.Variable 'net_pretrain/conv1_down/conv2d_9/kernel:0' shape=(3, 3, 4, 32) dtype=float32>, ...<tf.Variable 'net_pretrain_1/conv1_down/conv2d_15/kernel:0' shape=(3, 3, 4, 32) dtype=float32>, ...] 

我不确定问题出在哪里。太感谢了！

标签： pythontensorflow