tensorflow2.0+TF-lite 各种报错

generic_type: type "InterpreterWrapper" is already registered!

原因：tensorflow2.5.0rc0版本太高，
1.降低版本:

pip install tensorflow==2.3 

2.搜索整个代码库查找其他尝试注册"InterpreterWrapper"的地方，注释掉

3.可以尝试使用 clear_extension 函数删除注册表中的该类型

from google.protobuf.descriptor import FieldDescriptor
from tensorflow.lite.python.schema_py_generated import _TENSORFLOW_LITE_WRAPPERS_PROTOBUF_WRAPPER_DESCRIPTOR as descriptor

# Clear the extension registry for InterpreterWrapper
if 'InterpreterWrapper' in descriptor.extensions_by_name:
FieldDescriptor._extensions_by_name.pop('InterpreterWrapper', None)

 4.改用tf.lite

import tensorflow as tf

self.model = tf.lite.Interpreter(self.model_path + "/model.tflite")
self.model.allocate_tensors()
self.model_input = self.model.get_input_details()
self.model_output = self.model.get_output_details()
gpu_devices = tf.config.experimental.list_physical_devices('GPU')
if gpu_devices:
    # your_input_data 为输入数据
    self.model.set_tensor(self.model_input[0]['index'], your_input_data)
    self.model.invoke()
    output_tensor = self.model.get_tensor(self.model_output[0]['index'])

 

 

ValueErron: "batch_size" needs to be set for models with batch normalization layers.

模型搭建直接引用keras，前面去掉tensorflow

 

 

ImportError: cannot import name 'ParamSpec' from 'typing_extensions'

pip install typing-extensions==4.3.0

 或

pip install Twisted==22.4.0

 

 

 

cannot import name 'get_config' from 'tensorflow.python.eager.context'

keras版本太高

pip install keras==2.2.5

或者将keras都改成tensorflow.keras