InvalidArgumentError：在 loc 处需要可广播的形状（未知）

Question

背景

我对 Python 和机器学习完全陌生。我只是尝试根据我在互联网上找到的代码设置一个 UNet，并希望将其适应我正在一点一点地工作的情况。在尝试将 UNet .fit 到训练数据时，我收到以下错误：

InvalidArgumentError:  required broadcastable shapes at loc(unknown)
     [[node Equal (defined at <ipython-input-68-f1422c6f17bb>:1) ]] [Op:__inference_train_function_3847]

当我搜索它时，我得到了很多结果，但大多数都是不同的错误。

这是什么意思？而且，更重要的是，我该如何解决？

导致错误的代码

这个错误的上下文如下： 我想分割图像并标记不同的类。 我为训练、测试和验证数据设置了目录“trn”、“tst”和“val”。 dir_dat() 函数应用 os.path.join() 以获取相应 data set 的完整路径。 3 个文件夹中的每一个都有每个用整数标记的类的子目录。在每个文件夹中，都有一些 .tif 用于相应类的图像。

我定义了以下图像数据生成器（训练数据稀疏，因此进行了扩充）：

classes = np.array([ 0,  2,  4,  6,  8, 11, 16, 21, 29, 30, 38, 39, 51])
bs = 15 # batch size

augGen = ks.preprocessing.image.ImageDataGenerator(rotation_range = 365,
                                                   width_shift_range = 0.05,
                                                   height_shift_range = 0.05,
                                                   horizontal_flip = True,
                                                   vertical_flip = True,
                                                   fill_mode = "nearest") \
    .flow_from_directory(directory = dir_dat("trn"),
                         classes = [str(x) for x in classes.tolist()],
                         class_mode = "categorical",
                         batch_size = bs, seed = 42)
    
tst_batches = ks.preprocessing.image.ImageDataGenerator() \
    .flow_from_directory(directory = dir_dat("tst"),
                         classes = [str(x) for x in classes.tolist()],
                         class_mode = "categorical",
                         batch_size = bs, shuffle = False)

val_batches = ks.preprocessing.image.ImageDataGenerator() \
    .flow_from_directory(directory = dir_dat("val"),
                         classes = [str(x) for x in classes.tolist()],
                         class_mode = "categorical",
                         batch_size = bs)

然后我基于this example设置UNet。在这里，我更改了一些参数以使 UNet 适应这种情况（多类），即最后一层的激活和损失函数：

layer_in = ks.layers.Input(shape = (imgr, imgc, imgdim))
# convert pixel integer values to float
inVals = ks.layers.Lambda(lambda x: x / 255)(layer_in)

# Contraction path
c1 = ks.layers.Conv2D(16, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(inVals)
c1 = ks.layers.Dropout(0.1)(c1)
c1 = ks.layers.Conv2D(16, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c1)
p1 = ks.layers.MaxPooling2D((2, 2))(c1)

c2 = ks.layers.Conv2D(32, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(p1)
c2 = ks.layers.Dropout(0.1)(c2)
c2 = ks.layers.Conv2D(32, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c2)
p2 = ks.layers.MaxPooling2D((2, 2))(c2)
 
c3 = ks.layers.Conv2D(64, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(p2)
c3 = ks.layers.Dropout(0.2)(c3)
c3 = ks.layers.Conv2D(64, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c3)
p3 = ks.layers.MaxPooling2D((2, 2))(c3)
 
c4 = ks.layers.Conv2D(128, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(p3)
c4 = ks.layers.Dropout(0.2)(c4)
c4 = ks.layers.Conv2D(128, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c4)
p4 = ks.layers.MaxPooling2D(pool_size = (2, 2))(c4)
 
c5 = ks.layers.Conv2D(256, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(p4)
c5 = ks.layers.Dropout(0.3)(c5)
c5 = ks.layers.Conv2D(256, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c5)

# Expansive path 
u6 = ks.layers.Conv2DTranspose(128, (2, 2), strides = (2, 2), padding = "same")(c5)
u6 = ks.layers.concatenate([u6, c4])
c6 = ks.layers.Conv2D(128, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(u6)
c6 = ks.layers.Dropout(0.2)(c6)
c6 = ks.layers.Conv2D(128, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c6)
 
u7 = ks.layers.Conv2DTranspose(64, (2, 2), strides = (2, 2), padding = "same")(c6)
u7 = ks.layers.concatenate([u7, c3])
c7 = ks.layers.Conv2D(64, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(u7)
c7 = ks.layers.Dropout(0.2)(c7)
c7 = ks.layers.Conv2D(64, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c7)
 
u8 = ks.layers.Conv2DTranspose(32, (2, 2), strides = (2, 2), padding = "same")(c7)
u8 = ks.layers.concatenate([u8, c2])
c8 = ks.layers.Conv2D(32, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(u8)
c8 = ks.layers.Dropout(0.1)(c8)
c8 = ks.layers.Conv2D(32, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c8)
 
u9 = ks.layers.Conv2DTranspose(16, (2, 2), strides = (2, 2), padding = "same")(c8)
u9 = ks.layers.concatenate([u9, c1], axis = 3)
c9 = ks.layers.Conv2D(16, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(u9)
c9 = ks.layers.Dropout(0.1)(c9)
c9 = ks.layers.Conv2D(16, (3, 3), activation = "relu",
                            kernel_initializer = "he_normal", padding = "same")(c9)
 
out = ks.layers.Conv2D(1, (1, 1), activation = "softmax")(c9)
 
model = ks.Model(inputs = layer_in, outputs = out)
model.compile(optimizer = "adam", loss = "sparse_categorical_crossentropy", metrics = ["accuracy"])
model.summary()

最后，我定义了回调并运行了训练，这产生了错误：

cllbs = [
    ks.callbacks.EarlyStopping(patience = 4),
    ks.callbacks.ModelCheckpoint(dir_out("Checkpoint.h5"), save_best_only = True),
    ks.callbacks.TensorBoard(log_dir = './logs'),# log events for TensorBoard
    ]

model.fit(augGen, epochs = 5, validation_data = val_batches, callbacks = cllbs)

完整的控制台输出

这是运行最后一行时的完整输出（以防它有助于解决问题）：

trained = model.fit(augGen, epochs = 5, validation_data = val_batches, callbacks = cllbs)
Epoch 1/5
Traceback (most recent call last):

  File "<ipython-input-68-f1422c6f17bb>", line 1, in <module>
    trained = model.fit(augGen, epochs = 5, validation_data = val_batches, callbacks = cllbs)

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\keras\engine\training.py", line 1183, in fit
    tmp_logs = self.train_function(iterator)

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\def_function.py", line 889, in __call__
    result = self._call(*args, **kwds)

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\def_function.py", line 950, in _call
    return self._stateless_fn(*args, **kwds)

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\function.py", line 3023, in __call__
    return graph_function._call_flat(

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\function.py", line 1960, in _call_flat
    return self._build_call_outputs(self._inference_function.call(

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\function.py", line 591, in call
    outputs = execute.execute(

  File "c:\users\manuel\python\lib\site-packages\tensorflow\python\eager\execute.py", line 59, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,

InvalidArgumentError:  required broadcastable shapes at loc(unknown)
     [[node Equal (defined at <ipython-input-68-f1422c6f17bb>:1) ]] [Op:__inference_train_function_3847]

Function call stack:
train_function

Answer 1

尝试检查ks.layers.concatenate layers' 的输入是否具有相同的维度。例如ks.layers.concatenate([u7, c3])，这里检查 u7 和 c3 张量的形状是否相同，除了函数ks.layers.concatenate 的轴输入。 Axis = -1 默认，这是最后一个维度。为了说明你是否给出ks.layers.concatenate([u7,c3],axis=0)，那么除了u7和c3的第一个轴之外，所有其他轴的尺寸都应该完全匹配，例如u7.shape = [3,4,5], c3.shape = [6,4,5].

Answer 2

我遇到了同样的问题，因为我在模型（用于输出层）中使用了许多与标签/掩码数组中的实际类数不同的 n_classes。我看到你在这里有一个类似的问题：你有 13 个类，但你的输出层只有 1 个。最好的方法是避免硬编码类的数量，并且只在模型中传递一个变量（如 n_classes），然后在调用模型之前声明这个变量。例如 n_classes = y_Train.shape[-1] 或 n_classes = len(np.unique(y_Train))

Answer 3

当类标签的数量与输出层的输出形状的形状不匹配时，我遇到了这个问题。

例如，如果有 10 个类标签，我们将输出层定义为：

output = tf.keras.layers.Conv2D(5, (1, 1), activation = "softmax")(c9)

由于类标签 (10) 的数量不等于输出形状 (5)。 然后，我们会得到这个错误。

确保类标签的数量与输出层的输出形状相匹配。

Answer 4

我在这里发现了几个问题。该模型旨在用于具有多个类的语义分割（这就是为什么我将输出层激活更改为"softmax" 并设置"sparse_categorical_crossentropy" 损失）。因此，在 ImageDataGenerators 中，class_mode 必须设置为 None。不提供classes。相反，我需要将手动分类的图像插入为y。我猜初学者会犯很多初学者的错误。