导入Keras模型至DL4J后预测时出现形状不匹配错误求助

Keras训练模型代码 #

model = Sequential()
model.add(Embedding(50000, 128, input_length=10))
model.add(Conv1D(48, 5, activation='relu', padding='valid'))
model.add(GlobalMaxPooling1D())
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dropout(0.5))

model.add(Dense(7, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

DL4J预测Scala代码 #

val modelPath = File("C:/Users/Ayodele/Desktop/Development/classification.h5")
val model: MultiLayerNetwork = 
KerasModelImport.importKerasSequentialModelAndWeights(modelPath.absolutePath)
val list = arrayOf(1,2,3,4,5,6,7,8,9,10)
val inputs = 10
val features = Nd4j.create(1,inputs)
 for (i in 0 until inputs) {
        features.putScalar(intArrayOf(i), list.get(i))
    }
    System.out.println(features)
    val pred = model.predict(features) 

报错信息 #

Exception in thread "main" org.nd4j.linalg.exception.ND4JIllegalStateException: New 
shape length doesn't match original length: [288] vs [48]. Original shape: [1, 48] 
New Shape: [1, 288]
at org.nd4j.linalg.api.ndarray.BaseNDArray.reshape(BaseNDArray.java:3804)
at org.nd4j.linalg.api.ndarray.BaseNDArray.reshape(BaseNDArray.java:3749)
at org.nd4j.linalg.api.ndarray.BaseNDArray.reshape(BaseNDArray.java:3872)
at org.nd4j.linalg.api.ndarray.BaseNDArray.reshape(BaseNDArray.java:4099)
at 
org.deeplearning4j.preprocessors.KerasFlattenRnnPreprocessor.preProcess(KerasFlattenRnnPreprocessor.java:49)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.outputOfLayerDetached(MultiLayerNetwork.java:1299)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.output(MultiLayerNetwork.java:2467)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.output(MultiLayerNetwork.java:2430)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.output(MultiLayerNetwork.java:2421)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.output(MultiLayerNetwork.java:2408)
at org.deeplearning4j.nn.multilayer.MultiLayerNetwork.predict(MultiLayerNetwork.java:2270)

解决方案 #

• 问题根源：Keras中GlobalMaxPooling1D()输出的是(batch_size, filters)的2D张量（此处为(1,48)），后续添加的Flatten()属于冗余操作，不会改变张量形状，但DL4J的Keras模型导入器错误地将该Flatten()层识别为RNN类展平处理器，导致形状计算出现偏差，引发不匹配错误。

• 修复步骤：

  1. 修改Keras模型，移除冗余的Flatten()层，修改后的模型代码如下：

     model = Sequential()
     model.add(Embedding(50000, 128, input_length=10))
     model.add(Conv1D(48, 5, activation='relu', padding='valid'))
     model.add(GlobalMaxPooling1D())
     model.add(Dropout(0.5))
     # 移除冗余的Flatten层
     model.add(Dropout(0.5))
     
     model.add(Dense(7, activation='softmax'))
     model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
     

  2. 重新训练模型并导出为h5文件，再导入DL4J进行预测。
  3. 确认输入张量：DL4J中创建的features形状为(1,10)，与Keras模型的input_length=10匹配，无需调整；注意输入数据应为整数类型（Embedding层要求输入为词索引）。

内容的提问来源于stack exchange，提问作者HayWhy