将MobileNet模型从.h5转.tflite时遇属性缺失及LLVM类型推断错误求助

可行解决方案： #

1. 移除不必要的模型重新编译步骤 #

转TFLite不需要重新编译模型，这一步反而可能引发兼容性问题。直接删除loadmodel.compile(...)代码：

import tensorflow as tf

loadmodel = tf.keras.models.load_model('trained_model.h5')

converter = tf.lite.TFLiteConverter.from_keras_model(loadmodel) 
model = converter.convert()

with open('trained_model.tflite' , 'wb') as file: 
    file.write(model)

2. 强制使用TensorFlow内置Keras加载模型 #

Keras 3作为多后端框架，和TensorFlow 2.16.1存在适配问题。强制切换到TensorFlow原生Keras：

import tensorflow as tf
from tensorflow import keras

# 指定TensorFlow为Keras后端
keras.backend.set_backend('tensorflow')

loadmodel = tf.keras.models.load_model('trained_model.h5')

converter = tf.lite.TFLiteConverter.from_keras_model(loadmodel) 
model = converter.convert()

with open('trained_model.tflite' , 'wb') as file: 
    file.write(model)

3. 先转成SavedModel格式再转换TFLite #

.h5格式在跨版本/跨框架场景下易出问题，先转成TensorFlow原生的SavedModel格式：

import tensorflow as tf

# 加载模型并保存为SavedModel
loadmodel = tf.keras.models.load_model('trained_model.h5')
loadmodel.save('trained_savedmodel')

# 从SavedModel转换为TFLite
converter = tf.lite.TFLiteConverter.from_saved_model('trained_savedmodel')
model = converter.convert()

with open('trained_model.tflite' , 'wb') as file: 
    file.write(model)

4. 启用TF算子兼容模式 #

如果是算子不兼容导致的类型推断失败，开启TF算子支持：

import tensorflow as tf

loadmodel = tf.keras.models.load_model('trained_model.h5')

converter = tf.lite.TFLiteConverter.from_keras_model(loadmodel)
# 支持TFLite内置算子和TF原生算子
converter.target_spec.supported_ops = [
    tf.lite.OpsSet.TFLITE_BUILTINS,
    tf.lite.OpsSet.SELECT_TF_OPS
]
converter.allow_custom_ops = True

model = converter.convert()

with open('trained_model.tflite' , 'wb') as file: 
    file.write(model)

5. 指定输入输出数据类型 #

手动明确模型的输入输出类型为标准float32：

import tensorflow as tf

loadmodel = tf.keras.models.load_model('trained_model.h5')

converter = tf.lite.TFLiteConverter.from_keras_model(loadmodel)
converter.inference_input_type = tf.float32
converter.inference_output_type = tf.float32

model = converter.convert()

with open('trained_model.tflite' , 'wb') as file: 
    file.write(model)

内容的提问来源于stack exchange，提问作者Los