TFlite模型变换及使用

1. tflite模型转换

keras 模型转换

  1. import tensorflow as tf
  3. converter = tf.lite.TFLiteConverter.from_keras_model_file("keras_model.h5")
  4. tflite_model = converter.convert()
  5. open("converted_model.tflite", "wb").write(tflite_model)

SavedModel

  1. import tensorflow as tf
  3. converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
  4. tflite_model = converter.convert()
  5. open("converted_model.tflite", "wb").write(tflite_model)

PB：需要将meta和checkpoint通过freeze_graph输出为pb格式，再进行转化

  1. import tensorflow as tf
  3. graph_def_file = "/path/to/Downloads/mobilenet_v1_1.0_224/frozen_graph.pb"
  4. input_arrays = ["input"]
  5. output_arrays = ["MobilenetV1/Predictions/Softmax"]
  7. converter = tf.lite.TFLiteConverter.from_frozen_graph(
  8.   graph_def_file, input_arrays, output_arrays)
  9. tflite_model = converter.convert()
  10. open("converted_model.tflite", "wb").write(tflite_model)

2. tflite轻量化模型推理

导入包：

  1. import tflite_runtime.interpreter as tflite

这里使用tflite_runtime 库来实现模型的推理，该库独立与tensorflow可以单独运行，截至20210826官网上提供包在mac和linux上支持到py3.8，而win10下仅支持到python3.7。此外要成功打包需要对interpreter.py中函数进行修改。

  1. if not __file__.endswith(os.path.join('tflite_runtime', 'interpreter.py')):
  2.   # This file is part of tensorflow package.
  3.   # from tensorflow.python.util.lazy_loader import LazyLoader
  4.   # from tensorflow.python.util.tf_export import tf_export as _tf_export
  5.   # Lazy load since some of the performance benchmark skylark rules
  6.   # break dependencies. Must use double quotes to match code internal rewrite
  7.   # rule.
  8.   # pylint: disable=g-inconsistent-quotes
  9.   # _interpreter_wrapper = LazyLoader(
  10.   #     "_interpreter_wrapper", globals(),
  11.   #     "tensorflow.lite.python.interpreter_wrapper."
  12.   #     "tensorflow_wrap_interpreter_wrapper")
  13.   # pylint: enable=g-inconsistent-quotes

这里调用了tensorflow函数在程序打包后会提示找不到tensorflow依赖，修改后如下：

  1. from tflite_runtime import tensorflow_wrap_interpreter_wrapper as             _interpreter_wrapper
  3.   def _tf_export(*x, **kwargs):
  4.     del x, kwargs
  5.     return lambda x: x

本质上是将else中的部分在if下复制一份，让这两个导入的函数均来自tflite_runtime而不是tensorflow。

构建模型类如下：

  1. class ModelProcess:
  2.     def __init__(self,model_path=None) -> None:
  3.         self.interpreter=None 
  4.         self.result=None
  5.         self.__preProcess(model_path)
  7.     def __preProcess(self,path=None):
  8.         if path==None:
  9.             # 这里的操作是为了在将模型打包进程序后，能够从系统临时目录中加载模型，详细说明见pyinstaller --add-data
  10.             filename = self.resource_path(os.path.join("model_dir","full_model.tflite"))
  11.             self.interpreter=tflite.Interpreter(model_path=filename)
  12.         else:
  13.             self.interpreter=tflite.Interpreter(model_path=path)
  14.         self.interpreter.allocate_tensors()
  15.         # 解释器的输入和输出是由索引号来定位的
  16.         self.input_idx=self.interpreter.get_input_details()[0]['index']
  17.         self.output_idx=self.interpreter.get_output_details()[0]['index']
  19.     def compute(self,input_data):
  20.         """
  21.         模型计算，输入单条数据
  22.         """
  23.         self.interpreter.set_tensor(self.input_idx,input_data)
  24.         self.interpreter.invoke()
  25.         res=self.interpreter.get_tensor(self.output_idx)
  26.         res=np.squeeze(res)
  27.         return res 
  29.     def getResult(self,datas):
  30.         """
  31.         返回所有的计算结果
  32.         """
  33.         self.result=np.array([self.compute(data) for data in datas])
  35.     @staticmethod
  36.     def resource_path(relative_path):
  37.         if getattr(sys, 'frozen', False): #是否Bundle Resource
  38.             base_path = sys._MEIPASS
  39.         else:
  40.             base_path = os.path.abspath(".")
  41.         return os.path.join(base_path, relative_path)