TensorFlow Lite开发系列之python接口解析(一)

浅浅的花香味﹌ 2023-07-11 15:20 19阅读 0赞

> 环境: tensorflow2.x， 一定要使用linux系统，后期转换模型windows会出现bug

#### API解析 ####

网址：[https://tensorflow.google.cn/api\_docs/python/tf/lite][https_tensorflow.google.cn_api_docs_python_tf_lite]

<table> 
 <thead> 
  <tr> 
   <th>API</th> 
   <th>功能</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>class Interpreter</td> 
   <td>推理TensorFlow Lite模型</td> 
  </tr> 
  <tr> 
   <td>class TFLiteConverter</td> 
   <td>将TensorFlow模型转换为TensorFlow Lite模型</td> 
  </tr> 
  <tr> 
   <td>class Optimize</td> 
   <td>定义tflite图使用此API的优化</td> 
  </tr> 
  <tr> 
   <td>class OpsSet</td> 
   <td>TFLite模型的操作集</td> 
  </tr> 
 </tbody> 
</table>

常用的就这4个类，常用方法介绍：

# model_path：TF-Lite Flatbuffer文件的路径
    # model_content：模型的内容
    tf.lite.Interpreter(
        model_path=None, model_content=None, experimental_delegates=None
    )

allocate_tensors()   # 加载所有的tensor

get_input_details()  # 获取模型输入详细信息

get_output_details()  # 获取模型输出详细信息

# 获取输入张量的值（获取副本），该值可以从get_output_details中的“索引”字段中获得
    get_tensor( 
        tensor_index
    )

get_tensor_details()   # 返回值：包含张量信息的字典列表

invoke()  # 进行推理， 在调用它之前，请确保设置输入大小，分配张量和填充值

# tensor_index：要设置的张量的张量索引。该值可以从get_input_details中的“索引”字段中获得
    # value：要设置的张量值
    set_tensor(
        tensor_index, value
    )

# Converting a SavedModel to a TensorFlow Lite model.
    converter = lite.TFLiteConverter.from_saved_model(saved_model_dir)
    tflite_model = converter.convert()
    
    # Converting a tf.Keras model to a TensorFlow Lite model.
    converter = lite.TFLiteConverter.from_keras_model(model)
    tflite_model = converter.convert()
    
    # Converting ConcreteFunctions to a TensorFlow Lite model.
    converter = lite.TFLiteConverter.from_concrete_functions([func])
    tflite_model = converter.convert()

tf.lite.Optimize 
    类变量
    	DEFAULT
    	OPTIMIZE_FOR_LATENCY
    	OPTIMIZE_FOR_SIZE

### 完整例子： ###

#### 1. 训练一个模型用于后面转换TFLite模型与推理 ####

import tensorflow as tf
    from tensorflow import keras
    
    # 读取数据
    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
    x_train = tf.expand_dims(x_train, 3)
    y_train = keras.utils.to_categorical(y_train, num_classes=10)
    datasets = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    datasets = datasets.repeat(1).batch(10)
    
    # 定义模型
    img = keras.Input(shape=[28, 28, 1])
    
    x = keras.layers.Conv2D(filters=64, kernel_size=4, strides=1, padding='SAME', activation='relu')(img)
    x = keras.layers.AveragePooling2D(pool_size=2, strides=2, padding='SAME')(x)
    x = keras.layers.BatchNormalization()(x)
    x = keras.layers.Dropout(0.15)(x)
    
    x = keras.layers.Flatten()(x)
    x = keras.layers.Dense(512, activation='relu')(x)
    y_pred = keras.layers.Dense(10, activation='softmax')(x)
    
    model = keras.Model(inputs=img, outputs=y_pred)
    
    model.compile(optimizer=keras.optimizers.Adam(0.01),
                  loss=keras.losses.categorical_crossentropy,
                  metrics=['AUC', 'accuracy'])
    
    model.fit(datasets, epochs=1)
    
    model.save('./model.h5')

#### 2. 转换模型 ####

import tensorflow as tf
    from tensorflow import keras
    
    model = tf.keras.models.load_model("./model.h5")
    model.summary()
    
    converter = tf.lite.TFLiteConverter.from_keras_model(model)  # 生成转化器
    tflite_model = converter.convert()  # 进行转换
    open('./converted_model.tflite', 'wb').write(tflite_model)  # 写入

#### 3. 推理 ####

import tensorflow as tf
    from tensorflow import keras
    import numpy as np
    
    interpreter = tf.lite.Interpreter(model_path='./converted_model.tflite')  # 读入并生成interpreter
    interpreter.allocate_tensors()  # 加载所有的张量
    
    input_details = interpreter.get_input_details()  # 获取输入的信息
    output_details = interpreter.get_output_details()  # 获取输出的信息
    
    # 指定随机数进行预测
    input_shape = input_details[0]['shape']
    input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
    
    # 指定输入数据
    interpreter.set_tensor(input_details[0]['index'], input_data)
    
    # 调用模型进行推理
    interpreter.invoke()
    
    # 根据tensor索引取出推理的结果
    tflite_result = interpreter.get_tensor(output_details[0]['index'])

[https_tensorflow.google.cn_api_docs_python_tf_lite]: https://tensorflow.google.cn/api_docs/python/tf/lite