TensorFlow中Input has undefined rank错误的原因及修复方案

问题描述 #

自定义DenseNet模型时已明确指定输入形状，但运行时触发ValueError: Input has undefined rank. Received: input_shape=<unknown>.错误。

代码实现 #

def densenet(input_shape, n_classes, filters = 32):
    
    #batch norm + relu + conv
    def bn_rl_conv(x,filters,kernel=1,strides=1):
        
        x = BatchNormalization()(x)
        x = ReLU()(x)
        x = Conv2D(filters, kernel, strides=strides,padding = 'same')(x)
        return x
    
    def dense_block(x, repetition):
        
        for _ in range(repetition):
            y = bn_rl_conv(x, 4*filters)
            y = bn_rl_conv(y, filters, 3)
            x = concatenate([y,x])
        return x
        
    def transition_layer(x):
        
        x = bn_rl_conv(x, K.int_shape(x)[-1] //2 )
        x = AvgPool2D(2, strides = 2, padding = 'same')(x)
        return x
    
    inp = Input (input_shape)
    x = Conv2D(64, 7, strides = 2, padding = 'same')(inp)
    x = MaxPool2D(3, strides = 2, padding = 'same')(x)
                        
    for repetition in [2,4,6,4]:
        
        d = dense_block(x, repetition)
        x = transition_layer(d)
    x = GlobalAveragePooling2D()(x)
    
    output = Dense(n_classes, activation = 'softmax')(x)
    
    model = Model(inp, output)
    return model

input_shape = (1024,2,1)
num_classes = 24
model = densenet(input_shape,num_classes)

报错信息 #

/usr/local/lib/python3.7/dist-packages/keras/utils/traceback_utils.py in error_handler(*args, **kwargs)
     65     except Exception as e:  # pylint: disable=broad-except
     66       filtered_tb = _process_traceback_frames(e.__traceback__)
---> 67       raise e.with_traceback(filtered_tb) from None
     68     finally:
     69       del filtered_tb

/usr/local/lib/python3.7/dist-packages/keras/layers/normalization/batch_normalization.py in build(self, input_shape)
    296     if not input_shape.ndims:
    297       raise ValueError(
--> 298           f'Input has undefined rank. Received: input_shape={input_shape}.')
    299     ndims = len(input_shape)
    300 

ValueError: Input has undefined rank. Received: input_shape=<unknown>.

原因分析 #

错误核心在transition_layer函数中的K.int_shape(x)[-1] //2：

• 在Keras函数式API的模型构建阶段，张量x的静态形状尚未完全固化，K.int_shape(x)[-1]返回的是符号化张量而非具体整数。
• 当这个符号化值被传入bn_rl_conv作为Conv2D的filters参数时，Conv2D层无法确定输入维度（rank），进而导致后续BatchNormalization层触发输入rank未知的错误。

修复方案 #

将transition_layer中获取通道数的方式改为直接读取张量的静态形状属性x.shape[-1]，它能在模型构建阶段返回确定的整数值：

修改后的transition_layer函数：

def transition_layer(x):
    # 直接读取静态通道数并减半
    current_channels = x.shape[-1]
    x = bn_rl_conv(x, current_channels // 2)
    x = AvgPool2D(2, strides = 2, padding = 'same')(x)
    return x

若你的Keras版本对x.shape[-1]支持有限，也可改用tf.keras.backend.int_shape(x)[-1]，确保在构建时能解析为整数：

def transition_layer(x):
    current_channels = tf.keras.backend.int_shape(x)[-1]
    x = bn_rl_conv(x, current_channels // 2)
    x = AvgPool2D(2, strides = 2, padding = 'same')(x)
    return x

修改后重新运行代码，模型即可正常构建。

内容的提问来源于stack exchange，提问作者nechi