Variational AutoEncoder- Keras-logits 和标签必须具有相同的形状？

Question

我正在 Keras https://keras.io/examples/generative/vae/ 上浏览这个示例

但是，我正在尝试使用 200 x 200 尺寸的图像来复制它。我收到的具体错误是：

ValueError: logits and labels must have the same shape ((None, 8, 8, 1) vs (None, 200, 200, 3))

这是我的编码器，修改后的 keras.Input(shape=(200, 200, 3))

latent_dim = 2

encoder_inputs = keras.Input(shape=(200, 200, 3))
x = layers.Conv2D(32, 3, activation="relu", strides=2, padding="same")(encoder_inputs)
x = layers.Conv2D(64, 3, activation="relu", strides=2, padding="same")(x)
x = layers.Flatten()(x)
x = layers.Dense(16, activation="relu")(x)
z_mean = layers.Dense(latent_dim, name="z_mean")(x)
z_log_var = layers.Dense(latent_dim, name="z_log_var")(x)
z = Sampling()([z_mean, z_log_var])
encoder = keras.Model(encoder_inputs, [z_mean, z_log_var, z], name="encoder")
encoder.summary()

Model: "encoder"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_17 (InputLayer)           [(None, 200, 200, 3) 0                                            
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 100, 100, 32) 896         input_17[0][0]                   
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 50, 50, 64)   18496       conv2d_8[0][0]                   
__________________________________________________________________________________________________
flatten_4 (Flatten)             (None, 160000)       0           conv2d_9[0][0]                   
__________________________________________________________________________________________________
dense_16 (Dense)                (None, 16)           2560016     flatten_4[0][0]                  
__________________________________________________________________________________________________
z_mean (Dense)                  (None, 2)            34          dense_16[0][0]                   
__________________________________________________________________________________________________
z_log_var (Dense)               (None, 2)            34          dense_16[0][0]                   
__________________________________________________________________________________________________
sampling_3 (Sampling)           (None, 2)            0           z_mean[0][0]                     
                                                                 z_log_var[0][0]                  
==================================================================================================
Total params: 2,579,476
Trainable params: 2,579,476
Non-trainable params: 0

我相信错误出在我的解码器上，我试图将图层修改为 200 倍。

latent_inputs = keras.Input(shape=(latent_dim,))
x = layers.Dense(2 * 2 * 50, activation="relu")(latent_inputs)
x = layers.Reshape((2, 2, 50))(x) ##changed this
x = layers.Conv2DTranspose(64, 3, activation="relu", strides=2, padding="same")(x)
x = layers.Conv2DTranspose(32, 3, activation="relu", strides=2, padding="same")(x)
decoder_outputs = layers.Conv2DTranspose(1, 3, activation="sigmoid", padding="same")(x)
decoder = keras.Model(latent_inputs, decoder_outputs, name="decoder")
decoder.summary()

Model: "decoder"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_18 (InputLayer)        [(None, 2)]               0         
_________________________________________________________________
dense_17 (Dense)             (None, 200)               600       
_________________________________________________________________
reshape_12 (Reshape)         (None, 2, 2, 50)          0         
_________________________________________________________________
conv2d_transpose_13 (Conv2DT (None, 4, 4, 64)          28864     
_________________________________________________________________
conv2d_transpose_14 (Conv2DT (None, 8, 8, 32)          18464     
_________________________________________________________________
conv2d_transpose_15 (Conv2DT (None, 8, 8, 1)           289       
=================================================================
Total params: 48,217
Trainable params: 48,217
Non-trainable params: 0

我的 pic1 的图片尺寸是：

(312, 465)

然后我运行它，并遇到错误：

pic_1 = np.expand_dims(pic1, 0).astype("float32") / 255 

pic_1 = pic_1[:,-201:-1, 0:200] #Trim the picture to fit the input 200-by-200 dimensions

vae = VAE(encoder, decoder)
vae.compile(optimizer=keras.optimizers.Adam())
vae.fit(pic_1, epochs=30, batch_size=128)

我收到这个错误是什么原因：

        raise ValueError("logits and labels must have the same shape (%s vs %s)" %

    ValueError: logits and labels must have the same shape ((None, 8, 8, 1) vs (None, 200, 200, 3))

就像我说的，我认为我的解码器输出形状有问题，因为它与编码器的输入形状不匹配？任何帮助将不胜感激。

Answer 1

解码器不需要完全等于编码器（相同的层和长度），但需要输入与输出具有相同的形状。基本上，你可以改变解码器的层，但最终的层形状必须是（200,200,3）。简单的方法是根据编码器镜像解码器。举个例子：

latent_inputs = keras.Input(shape=(latent_dim,))
x = layers.Dense(16, activation="relu")(latent_inputs)
x = layers.Dense(50 * 50 * 64, activation="relu")(x)
x = layers.Reshape((50, 50, 64))(x)
x = layers.Conv2DTranspose(32, 3, activation="relu", strides=2, padding="same")(x)
decoder_outputs = layers.Conv2DTranspose(3, 3, activation="sigmoid", strides=2, padding="same")(x)
decoder = keras.Model(latent_inputs, decoder_outputs, name="decoder")
decoder.summary()

解码器与编码器不同但输出形状始终为 (200,200,3) 的其他示例是：

latent_inputs = keras.Input(shape=(latent_dim,))
x = layers.Dense(8, activation="relu")(latent_inputs)
x = layers.Dense(16, activation="relu")(x)
x = layers.Dense(50 * 50 * 64, activation="relu")(x)
x = layers.Reshape((50, 50, 64))(x)
x = layers.Conv2DTranspose(64, 3, activation="relu", strides=1, padding="same")(x)
x = layers.Conv2DTranspose(32, 3, activation="relu", strides=1, padding="same")(x)
x = layers.Conv2DTranspose(16, 3, activation="relu", strides=2, padding="same")(x)
decoder_outputs = layers.Conv2DTranspose(3, 3, activation="sigmoid", strides=2, padding="same")(x)
decoder = keras.Model(latent_inputs, decoder_outputs, name="decoder")
decoder.summary()