Question

我是一名本科生，致力于使用TensorFlow重新创建论文“使用生成的对抗网络实现逼真的单图像超分辨率”。训练完成后尝试测试网络时，出现关于不兼容形状的错误（即使它适用于训练）。看来此错误与Keras Sequential有关。

错误结尾给出：

File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [24,24,64] vs. [16,256,256,64] [Op:Mul] name: generator/sequential/p_re_lu/mul/

其中训练输入形状为(16,24,24,3)，测试输入形状为(1,256,256,3)。网络中的第一层是tf.keras.layers.Conv2D(64, 9, strides=1, padding='same'),层。

我从顺序中分离出了一些层，并打印了输出形状作为验证。我留在顺序块中的图层会抛出相同的错误，而分开的图层可以正常工作

输出：

x_in: (16, 24, 24, 3)
x1: (16, 24, 24, 64)
x2a: (16, 24, 24, 64)
x2: (16, 24, 24, 64)
x3: (16, 24, 24, 64)
x4: (16, 48, 48, 256)
x5: (16, 96, 96, 256)
x_out: (16, 96, 96, 3)
x_in: (1, 256, 256, 3)
x1: (1, 256, 256, 64)
x2a: (1, 256, 256, 64)
Traceback (most recent call last):
  File "srgan.py", line 249, in <module> ...

顺序块：

        self.residual = tf.keras.Sequential([
            Residual(cfg, num_filters),
            Residual(cfg, num_filters),
            Residual(cfg, num_filters),
            Residual(cfg, num_filters),
        ])

通话功能

def call(self, x_in):

        print(f'x_in: {x_in.shape}')

        x1a = tf.keras.layers.Conv2D(64, 9, strides=1, padding='same')(x_in)
        x1 = tf.keras.layers.PReLU()(x1a)
        #x1 = self.start(x_in)
        print(f'x1: {x1.shape}')   
        x2a = Residual(self.cfg, 64)(x1) 
        print(f'x2a: {x2a.shape}')   
        x2 = self.residual(x1)

我希望序列与分离的层相同。我是否缺少某些东西或这是一个错误？

edit1：当我说我分离层时，我的意思是将层移到顺序之外。

我将它们全部删除，然后再次尝试模型。该错误现在针对tf.keras.layers.PReLU()，这是有道理的，因为该错误提到了name: generator/sequential/p_re_lu/mul/

很奇怪，这样会给比较形状带来错误

这是模型摘要

Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              multiple                  15616     
_________________________________________________________________
p_re_lu (PReLU)              multiple                  36864     
_________________________________________________________________
residual (Residual)          multiple                  111232    
_________________________________________________________________
residual_1 (Residual)        multiple                  111232    
_________________________________________________________________
residual_2 (Residual)        multiple                  111232    
_________________________________________________________________
residual_3 (Residual)        multiple                  111232    
_________________________________________________________________
conv2d_9 (Conv2D)            multiple                  331840    
_________________________________________________________________
batch_normalization_v2_8 (Ba multiple                  256       
_________________________________________________________________
conv2d_transpose (Conv2DTran multiple                  147712    
_________________________________________________________________
p_re_lu_5 (PReLU)            multiple                  589824    
_________________________________________________________________
conv2d_transpose_1 (Conv2DTr multiple                  590080    
_________________________________________________________________
p_re_lu_6 (PReLU)            multiple                  2359296   
_________________________________________________________________
conv2d_10 (Conv2D)           multiple                  6915      
=================================================================
Total params: 4,523,331
Trainable params: 4,522,179
Non-trainable params: 1,152

Answer 1

事实证明，您需要重新初始化以下模型。 Is loading in eager TensorFlow broken right now?

有多种方法可以做到，如上面链接中问题的答案所述。

仅当使用Keras Sequential

1 个答案: