将keras模型转换为core ml模型时出错

如何解决将keras模型转换为core ml模型时出错

我一直在尝试将 Keras 模型转换为核心 mlmodel。但是，在这样做的同时，我得到了多个输入和输出。我只想要以下作为输入和输出。

name: "input"
type {  
    imageType {    
        width: 256    
        height: 64    
        colorSpace: GRAYSCALE  }
}

name: "output"
type {
  dictionaryType {
    stringKeyType {
    }
  }
}

我得到的当前模型。

type {
  imageType {
    width: 256
    height: 64
    colorSpace: GRAYSCALE
  }
},name: "lstm1_h_in"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm1_c_in"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm1_h_in_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm1_c_in_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm2_h_in"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm2_c_in"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm2_h_in_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
},name: "lstm2_c_in_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
  isOptional: true
}
]

[name: "output"
type {
  dictionaryType {
    stringKeyType {
    }
  }
},name: "lstm1_h_out"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm1_c_out"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm1_h_out_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm1_c_out_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm2_h_out"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm2_c_out"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm2_h_out_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "lstm2_c_out_rev"
type {
  multiArrayType {
    shape: 256
    dataType: DOUBLE
  }
},name: "classLabel"
type {
  stringType {
  }
}
]

我正在使用此代码进行转换：

from keras.models import load_model
import coremltools as ct

class_labels = ["0","1","2","3","4","5","6","7","8","9","A","B","C","D","E","F","G","H","I","J","K","L","M","N","O","P","Q","R","S","T","U","V","W","X","Y","Z","a","b","c","d","e","f","g","h","i","j","k","l","m","n","o","p","q","r","s","t","u","v","w","x","y","z","-"]

tf_keras_model = load_model("HTRAdam.h5")
mlmodel = ct.converters.keras.convert('HTRAdam.h5',input_names=['image'],output_names=['output'],class_labels=class_labels,image_input_names='image')
mlmodel.save('HTR.mlmodel')

当我在此输入图像时，我没有得到任何输出。我的意思是图像输出为零。 我进入调试模式，但在模型预测中找不到任何错误。

我猜问题可能出在编译模型上，但我不知道如何解决。 感谢您的帮助。

解决方法

如果您的模型包含 LSTM 层，Core ML 期望您手动传入 LSTM 状态，这就是您获得所有这些额外输入和输出的原因。 Keras 会为您管理此状态，但 Core ML 不会。