PointNet训练fit函数报错：标签与Logits维度不匹配

问题描述 #

训练PointNet模型时，运行model.fit出现如下错误：

Exception has occurred: ValueError
`labels.shape` must equal `logits.shape` except for the last dimension. Received: labels.shape=(300000,) and logits.shape=(60, 2)

模型实现代码 #

def conv_bn(x, filters):
    x = keras.layers.Conv1D(filters, kernel_size=1, padding="valid")(x)
    x = keras.layers.BatchNormalization(momentum=0.0)(x)
    return keras.layers.Activation("relu")(x)


def dense_bn(x, filters):
    x = keras.layers.Dense(filters)(x)
    x = keras.layers.BatchNormalization(momentum=0.0)(x)
    return keras.layers.Activation("relu")(x)

class OrthogonalRegularizer(keras.regularizers.Regularizer):
    def __init__(self, num_features, l2reg=0.001):
        self.num_features = num_features
        self.l2reg = l2reg
        self.eye = tf.eye(num_features)

    def __call__(self, x):
        x = tf.reshape(x, (-1, self.num_features, self.num_features))
        xxt = tf.tensordot(x, x, axes=(2, 2))
        xxt = tf.reshape(xxt, (-1, self.num_features, self.num_features))
        return tf.reduce_sum(self.l2reg * tf.square(xxt - self.eye))
def tnet(inputs, num_features):

    # Initalise bias as the indentity matrix
    bias = keras.initializers.Constant(np.eye(num_features).flatten())
    reg = OrthogonalRegularizer(num_features)

    x = conv_bn(inputs, 32)
    x = conv_bn(x, 64)
    x = conv_bn(x, 512)
    x = keras.layers.GlobalMaxPooling1D()(x)
    x = dense_bn(x, 256)
    x = dense_bn(x, 128)
    x = keras.layers.Dense(num_features * num_features,
        kernel_initializer="zeros",bias_initializer=bias,
        activity_regularizer=reg,)(x)
    feat_T = keras.layers.Reshape((num_features, num_features))(x)
    # Apply affine transformation to input features
    return keras.layers.Dot(axes=(2, 1))([inputs, feat_T])
inputs = keras.Input(shape=(BATCH_SIZE,7))
print(inputs)
x = tnet(inputs, 7)
x = conv_bn(x, 32)
x = conv_bn(x, 32)
x = tnet(x, 32)
x = conv_bn(x, 32)
x = conv_bn(x, 64)
x = conv_bn(x, 512)
x = keras.layers.GlobalMaxPooling1D()(x)
x = dense_bn(x, 256)
x = keras.layers.Dropout(0.3)(x)
x = dense_bn(x, 128)
x = keras.layers.Dropout(0.3)(x)
outputs = keras.layers.Dense(NUM_CLASSES, activation="softmax")(x)
model = keras.Model(inputs=inputs, outputs=outputs, name="pointnet")
model.summary()
model.compile(loss="sparse_categorical_crossentropy",
    optimizer=keras.optimizers.Adam(learning_rate=0.001),
    metrics=["sparse_categorical_accuracy"],)
model.fit(train_dataset, epochs=12, validation_data=val_dataset)

模型结构摘要 #

__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to
==================================================================================================     
 input_1 (InputLayer)           [(None, 5000, 7)]    0           []

 conv1d (Conv1D)                (None, 5000, 32)     256         ['input_1[0][0]']

 batch_normalization (BatchNorm  (None, 5000, 32)    128         ['conv1d[0][0]']
 alization)

 activation (Activation)        (None, 5000, 32)     0           ['batch_normalization[0][0]']

 conv1d_1 (Conv1D)              (None, 5000, 64)     2112        ['activation[0][0]']

 batch_normalization_1 (BatchNo  (None, 5000, 64)    256         ['conv1d_1[0][0]']
 rmalization)

 activation_1 (Activation)      (None, 5000, 64)     0           ['batch_normalization_1[0][0]']       

 conv1d_2 (Conv1D)              (None, 5000, 512)    33280       ['activation_1[0][0]']

 batch_normalization_2 (BatchNo  (None, 5000, 512)   2048        ['conv1d_2[0][0]']
 rmalization)

 activation_2 (Activation)      (None, 5000, 512)    0           ['batch_normalization_2[0][0]']       

 global_max_pooling1d (GlobalMa  (None, 512)         0           ['activation_2[0][0]']
 xPooling1D)

 dense (Dense)                  (None, 256)          131328      ['global_max_pooling1d[0][0]']       

 batch_normalization_3 (BatchNo  (None, 256)         1024        ['dense[0][0]']
 rmalization)

 activation_3 (Activation)      (None, 256)          0           ['batch_normalization_3[0][0]']       

 dense_1 (Dense)                (None, 128)          32896       ['activation_3[0][0]']

 batch_normalization_4 (BatchNo  (None, 128)         512         ['dense_1[0][0]']
 rmalization)

 activation_4 (Activation)      (None, 128)          0           ['batch_normalization_4[0][0]']       

 dense_2 (Dense)                (None, 49)           6321        ['activation_4[0][0]']

 reshape (Reshape)              (None, 7, 7)         0           ['dense_2[0][0]']

 dot (Dot)                      (None, 5000, 7)      0           ['input_1[0][0]',
                                                                   'reshape[0][0]']

 conv1d_3 (Conv1D)              (None, 5000, 32)     256         ['dot[0][0]']

 batch_normalization_5 (BatchNo  (None, 5000, 32)    128         ['conv1d_3[0][0]']
 rmalization)

 activation_5 (Activation)      (None, 5000, 32)     0           ['batch_normalization_5[0][0]']       

 conv1d_4 (Conv1D)              (None, 5000, 32)     1056        ['activation_5[0][0]']

 batch_normalization_6 (BatchNo  (None, 5000, 32)    128         ['conv1d_4[0][0]']
 rmalization)

 activation_6 (Activation)      (None, 5000, 32)     0           ['batch_normalization_6[0][0]']       

 conv1d_5 (Conv1D)              (None, 5000, 32)     1056        ['activation_6[0][0]']

 batch_normalization_7 (BatchNo  (None, 5000, 32)    128         ['conv1d_5[0][0]']
 rmalization)

 activation_7 (Activation)      (None, 5000, 32)     0           ['batch_normalization_7[0][0]']       

 conv1d_6 (Conv1D)              (None, 5000, 64)     2112        ['activation_7[0][0]']

 batch_normalization_8 (BatchNo  (None, 5000, 64)    256         ['conv1d_6[0][0]']
 rmalization)

 activation_8 (Activation)      (None, 5000, 64)     0           ['batch_normalization_8[0][0]']       

 conv1d_7 (Conv1D)              (None, 5000, 512)    33280       ['activation_8[0][0]']

 batch_normalization_9 (BatchNo  (None, 5000, 512)   2048        ['conv1d_7[0][0]']
 rmalization)

 activation_9 (Activation)      (None, 5000, 512)    0           ['batch_normalization_9[0][0]']       

 global_max_pooling1d_1 (Global  (None, 512)         0           ['activation_9[0][0]']
 MaxPooling1D)

 dense_3 (Dense)                (None, 256)          131328      ['global_max_pooling1d_1[0][0]']      

 batch_normalization_10 (BatchN  (None, 256)         1024        ['dense_3[0][0]']
 ormalization)

 activation_10 (Activation)     (None, 256)          0           ['batch_normalization_10[0][0]']      

 dense_4 (Dense)                (None, 128)          32896       ['activation_10[0][0]']

 batch_normalization_11 (BatchN  (None, 128)         512         ['dense_4[0][0]']
 ormalization)

 activation_11 (Activation)     (None, 128)          0           ['batch_normalization_11[0][0]']      

 dense_5 (Dense)                (None, 1024)         132096      ['activation_11[0][0]']

 reshape_1 (Reshape)            (None, 32, 32)       0           ['dense_5[0][0]']

 dot_1 (Dot)                    (None, 5000, 32)     0           ['activation_6[0][0]',
                                                                   'reshape_1[0][0]']

 conv1d_8 (Conv1D)              (None, 5000, 32)     1056        ['dot_1[0][0]']

 batch_normalization_12 (BatchN  (None, 5000, 32)    128         ['conv1d_8[0][0]']
 ormalization)

 activation_12 (Activation)     (None, 5000, 32)     0           ['batch_normalization_12[0][0]']      

 conv1d_9 (Conv1D)              (None, 5000, 64)     2112        ['activation_12[0][0]']

 batch_normalization_13 (BatchN  (None, 5000, 64)    256         ['conv1d_9[0][0]']
 ormalization)

 activation_13 (Activation)     (None, 5000, 64)     0           ['batch_normalization_13[0][0]']      

 conv1d_10 (Conv1D)             (None, 5000, 512)    33280       ['activation_13[0][0]']

 batch_normalization_14 (BatchN  (None, 5000, 512)   2048        ['conv1d_10[0][0]']
 ormalization)

 activation_14 (Activation)     (None, 5000, 512)    0           ['batch_normalization_14[0][0]']      

 global_max_pooling1d_2 (Global  (None, 512)         0           ['activation_14[0][0]']
 MaxPooling1D)

 dense_6 (Dense)                (None, 256)          131328      ['global_max_pooling1d_2[0][0]']      

 batch_normalization_15 (BatchN  (None, 256)         1024        ['dense_6[0][0]']
 ormalization)

 activation_15 (Activation)     (None, 256)          0           ['batch_normalization_15[0][0]']      

 dropout (Dropout)              (None, 256)          0           ['activation_15[0][0]']

 dense_7 (Dense)                (None, 128)          32896       ['dropout[0][0]']

 batch_normalization_16 (BatchN  (None, 128)         512         ['dense_7[0][0]']
 ormalization)

 activation_16 (Activation)     (None, 128)          0           ['batch_normalization_16[0][0]']      

 dropout_1 (Dropout)            (None, 128)          0           ['activation_16[0][0]']

 dense_8 (Dense)                (None, 2)            258         ['dropout_1[0][0]']

==================================================================================================      
Total params: 753,363
Trainable params: 747,283
Non-trainable params: 6,080
__________________________________________________________________________________________________      

错误分析 #

从错误信息和模型结构可定位两个核心问题：

1. 输入层定义错误：Keras的Input层shape参数不应包含批量大小（BATCH_SIZE），批量维度由训练数据集自动处理，只需指定单个样本的形状。你的模型摘要中输入shape为(None,5000,7)，说明BATCH_SIZE变量值为5000，这会导致模型将点云点数误判为批量大小。
2. 标签维度与任务不匹配：当前模型是点云级分类结构（最后使用GlobalMaxPooling1D，输出shape为(None,2)，对应每个点云样本1个分类结果），但标签shape为(300000,)，相当于每个点对应1个标签（60个批量样本×5000个点=300000），两者维度完全不匹配。

解决方案 #

方案1：点云分类任务（每个点云对应1个标签） #

1. 修正输入层定义：
   将输入层代码修改为：

   # NUM_POINTS为每个点云的点数（如5000），BATCH_SIZE无需写在这里
   inputs = keras.Input(shape=(NUM_POINTS,7))
   

2. 调整数据集标签：
   确保train_dataset和val_dataset返回的标签为每个点云对应1个值，即标签shape为(BATCH_SIZE,)。若原标签是每个点的标注，需将每个点云的所有点标签合并为一个（如取众数，或确认数据集本身的点云级标签）。

方案2：点分割任务（每个点对应1个标签） #

若任务是对每个点进行分类（如语义分割），需修改模型结构以匹配点级输出：

1. 移除最后的GlobalMaxPooling1D：
   删除x = keras.layers.GlobalMaxPooling1D()(x)这一行，保留特征的(None,5000,512)形状。
2. 调整输出层：
   使用Conv1D替代Dense，让输出对应每个点的分类结果，确保输出shape为(None,5000,NUM_CLASSES)：

   outputs = keras.layers.Conv1D(NUM_CLAS