我正在尝试构建一个变分自动编码器。我在运行 model.fit 时遇到错误，我不明白

Question

纪元 1/10

TypeError Traceback（最后一次调用） 在() 1 # 火车模型 ----> 2 model.fit(X_train, y_train, epochs=10, batch_size=16, validation_data=(X_val, y_val))

1帧 /usr/local/lib/python3.9/dist-packages/keras/engine/training.py 中的 tf__train_function(迭代器) 13 尝试： 14 do_return = 真 ---> 15 retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), 无, fscope) 16 除了： 17 do_return = 假

TypeError：在用户代码中：

File "/usr/local/lib/python3.9/dist-packages/keras/engine/training.py", line 1284, in train_function  *
    return step_function(self, iterator)
File "/usr/local/lib/python3.9/dist-packages/keras/engine/training.py", line 1268, in step_function  **
    outputs = model.distribute_strategy.run(run_step, args=(data,))
File "/usr/local/lib/python3.9/dist-packages/keras/engine/training.py", line 1249, in run_step  **
    outputs = model.train_step(data)
File "/usr/local/lib/python3.9/dist-packages/keras/engine/training.py", line 1051, in train_step
    loss = self.compute_loss(x, y, y_pred, sample_weight)
File "/usr/local/lib/python3.9/dist-packages/keras/engine/training.py", line 1109, in compute_loss
    return self.compiled_loss(
File "/usr/local/lib/python3.9/dist-packages/keras/engine/compile_utils.py", line 317, in __call__
    self._total_loss_mean.update_state(
File "/usr/local/lib/python3.9/dist-packages/keras/utils/metrics_utils.py", line 77, in decorated
    update_op = update_state_fn(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/keras/metrics/base_metric.py", line 140, in update_state_fn
    return ag_update_state(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/keras/metrics/base_metric.py", line 477, in update_state  **
    sample_weight = tf.__internal__.ops.broadcast_weights(
File "/usr/local/lib/python3.9/dist-packages/keras/engine/keras_tensor.py", line 283, in __array__
    raise TypeError(

TypeError: You are passing KerasTensor(type_spec=TensorSpec(shape=(), dtype=tf.float32, name=None), name='Placeholder:0', description="created by layer 'tf.cast_2'"), an intermediate Keras symbolic input/output, to a TF API that does not allow registering custom dispatchers, such as `tf.cond`, `tf.function`, gradient tapes, or `tf.map_fn`. Keras Functional model construction only supports TF API calls that *do* support dispatching, such as `tf.math.add` or `tf.reshape`. Other APIs cannot be called directly on symbolic Kerasinputs/outputs. You can work around this limitation by putting the operation in a custom Keras layer `call` and calling that layer on this symbolic input/output.

这是代码

    # Set path to directory containing images
    path_to_images = '/content/drive/MyDrive/Heatsource504'

    # Define regular expression pattern to extract coordinates
    pattern = r'x_(\d+)y_(\d+)\.jpg'

    # Initialize empty lists to store image data and corresponding coordinates
    images = []
    heatmaps = []
    coordinates = []

    # Loop through images and extract data and coordinates
    for filename in os.listdir(path_to_images):
        match = re.search(pattern, filename)
        if match:
            x_coord = int(match.group(1))
            y_coord = int(match.group(2))
            img = Image.open(os.path.join(path_to_images, filename))
            img = img.resize((200, 200))  # resize image to desired dimensions
            img_array = np.array(img)  # convert image to numpy array

            # Convert image to grayscale if it is not already grayscale
            if len(img_array.shape) == 3 and img_array.shape[2] == 3:
                img_array = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)

            # Check if image is of type uint8, and convert it to uint8 if necessary
            if img_array.dtype != 'uint8':
                img_array = img_array.astype('uint8')

            # Normalize pixel values of input image to be between 0 and 1
            img_array = img_array / 255.0

            # Add input image and corresponding output image to lists
            heatmaps.append(img_array)
            images.append(cv2.imread(os.path.join(path_to_images, filename)))
            coordinates.append([x_coord, y_coord])

    # Convert lists to numpy arrays
    images = np.array(images)
    heatmaps = np.array(heatmaps)
    coordinates = np.array(coordinates)

    # Split data into training and validation sets
    X_train, X_val, y_train, y_val, coords_train, coords_val = train_test_split(heatmaps, images, coordinates, test_size=0.2, random_state=42)

    return X_train, y_train, coords_train, X_val, y_val, coords_val


input_shape = (200, 200, 1)

# Encoder
inputs = Input(shape=input_shape)
x = Conv2D(filters=16, kernel_size=3, padding='valid', activation='relu')(inputs)
x = Conv2D(filters=32, kernel_size=3, padding='valid', activation='relu')(x)
x = Conv2D(filters=64, kernel_size=3, padding='valid', activation='relu')(x)
x = Conv2D(filters=128, kernel_size=3, padding='valid', activation='relu')(x)
x = Conv2D(filters=256, kernel_size=3, padding='same', activation='relu')(x)
x = Flatten()(x)
x = Dense(units=128, activation='relu')(x)

# Define latent variables
latent_dim = 10
mu = Dense(units=latent_dim)(x)
log_var = Dense(units=latent_dim)(x)

# Reparameterization trick
def sampling(args):
    mu, log_var = args
    epsilon = K.random_normal(shape=K.shape(mu))
    return mu + K.exp(log_var / 2) * epsilon
# Sample latent variables
z = Lambda(sampling)([mu, log_var])

# Decoder
x = Dense(units=128, activation='relu')(z)
x = Dense(units=8 * 8 * 128, activation='relu')(x)
x = Reshape(target_shape=(8, 8, 128))(x)
x = Conv2DTranspose(filters=128, kernel_size=3, padding='same', activation='relu')(x)
x = Conv2DTranspose(filters=64, kernel_size=4, padding='valid', activation='relu')(x)
x = Conv2DTranspose(filters=32, kernel_size=2, padding='valid', activation='relu')(x)
x = Conv2DTranspose(filters=16, kernel_size=3, padding='valid', activation='relu')(x)
x = Conv2DTranspose(filters=1, kernel_size=2, padding='valid', activation='sigmoid')(x)

# Define VAE loss function
def vae_loss(y_true, y_pred):
    reconstruction_loss = binary_crossentropy(K.flatten(y_true), K.flatten(y_pred))
    reconstruction_loss *= input_shape[0] * input_shape[1] * input_shape[2]
    kl_loss = 1 + log_var - K.square(mu) - K.exp(log_var)
    kl_loss = K.sum(kl_loss, axis=-1)
    kl_loss *= -0.5
    vae_loss = K.mean(reconstruction_loss + kl_loss)
    return vae_loss

model = Model(inputs=inputs, outputs=x)
model.compile(optimizer='adam', loss=vae_loss)

# Load data
X_train, y_train, coords_train, X_val, y_val, coords_val = load_data4()
model.fit(X_train, y_train, epochs=10, batch_size=16, validation_data=(X_val, y_val))```

Answer 1

我记得这件事也发生在我身上。 tensorflow 似乎不再支持这样的 vae_loss 函数了。
我有 2 个解决方案，我将在此处粘贴简短的解决方案。您需要像这样添加损失，而不是创建 vae_loss 函数：

reconstruction_loss = binary_crossentropy(K.flatten(inputs), K.flatten(outputs)) #I think in your case outputs is x
reconstruction_loss *= input_shape[0] * input_shape[1] * input_shape[2]
kl_loss = 1 + log_var - K.square(mu) - K.exp(log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5
vae_loss = K.mean(reconstruction_loss + kl_loss)

model = Model(inputs=inputs, outputs=x)
model.add_loss(vae_loss) #This is the key

model.compile(optimizer='adam', loss=vae_loss)

这个问题我还有另一种解决方案，那就是使用自定义模型。你可以在这里找到这个解决方案：
https://keras.io/examples/generative/vae/