import tensorflow as tf kmodel = tf.keras.models.load_model('model.h5') kmodel.save('model.kmodle')出现WARNING:absl:Found untraced functions such as _jit_compiled_convolution_op, _jit_compiled_convolution_op, _jit_compiled_convolution_op, _jit_compiled_convolution_op, _jit_compiled_convolution_op while saving (showing 5 of 7). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: model.kmodle\assets INFO:tensorflow:Assets written to: model.kmodle\assets的警告怎么修改解决

这个警告是因为在保存 Keras 模型时，某些自定义的 Op 操作没有被跟踪到。这将导致在加载模型时，这些 Op 操作将不可直接调用。

你可以尝试使用 `tf.keras.experimental.export_saved_model` 函数来保存模型，例如：

import tensorflow as tf

kmodel = tf.keras.models.load_model('model.h5')
tf.keras.experimental.export_saved_model(kmodel, 'model_kmodle')

这将使用 TensorFlow SavedModel 格式保存模型，可以避免这个警告。

相关问题

.tflite转k210的.kmodel模型代码

将.tflite模型转换为.kmodel模型的代码如下：

import tensorflow as tf
import numpy as np
from tensorflow.keras import Model
from tensorflow.keras.layers import Input

# 加载tflite模型
interpreter = tf.lite.Interpreter(model_path="model.tflite")
interpreter.allocate_tensors()

# 获取输入输出张量的索引
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# 创建Keras模型
inputs = Input(shape=input_details[0]['shape'][1:])
x = tf.keras.layers.Lambda(lambda x: x / 255.)(inputs) # 对输入进行归一化
for i, layer in enumerate(interpreter.get_tensor_details()):
    if i < 2:
        continue
    op = layer['op']
    name = layer['name']
    shape = layer['shape']
    weights = interpreter.get_tensor(layer['index'])
    if op == 'CONV_2D':
        x = tf.keras.layers.Conv2D(
            filters=shape[-1],
            kernel_size=shape[1:3],
            strides=layer['stride'][1:3],
            padding='same',
            use_bias=layer['quantization_parameters']['quantized_dimension'] == 3,
            kernel_initializer=tf.keras.initializers.Constant(weights[0]),
            bias_initializer=tf.keras.initializers.Constant(weights[1])
        )(x)
    elif op == 'DEPTHWISE_CONV_2D':
        x = tf.keras.layers.DepthwiseConv2D(
            kernel_size=shape[1:3],
            strides=layer['stride'][1:3],
            padding='same',
            use_bias=layer['quantization_parameters']['quantized_dimension'] == 3,
            depthwise_initializer=tf.keras.initializers.Constant(weights[0]),
            bias_initializer=tf.keras.initializers.Constant(weights[1])
        )(x)
    elif op == 'AVERAGE_POOL_2D':
        x = tf.keras.layers.AveragePooling2D(
            pool_size=shape[1:3],
            strides=layer['stride'][1:3],
            padding='same'
        )(x)
    elif op == 'MAX_POOL_2D':
        x = tf.keras.layers.MaxPooling2D(
            pool_size=shape[1:3],
            strides=layer['stride'][1:3],
            padding='same'
        )(x)
    elif op == 'ADD':
        x = tf.keras.layers.Add()([x, tf.keras.layers.Lambda(lambda y: y[..., 0])(x)])
    elif op == 'RELU':
        x = tf.keras.layers.ReLU()(x)
    elif op == 'RESHAPE':
        x = tf.keras.layers.Reshape(target_shape=shape[1:])(x)
    elif op == 'FULLY_CONNECTED':
        x = tf.keras.layers.Dense(
            units=shape[-1],
            use_bias=True,
            kernel_initializer=tf.keras.initializers.Constant(weights[0]),
            bias_initializer=tf.keras.initializers.Constant(weights[1])
        )(x)
    elif op == 'SOFTMAX':
        x = tf.keras.layers.Softmax()(x)
outputs = x
keras_model = Model(inputs, outputs)

# 将Keras模型转换为K210的.kmodel模型
import tensorflow.keras.backend as K
from tensorflow.keras.models import model_from_json

# 保存Keras模型的权重
keras_model.save_weights('weights.h5')

# 保存Keras模型的结构
keras_model_json = keras_model.to_json()
with open('model.json', 'w') as f:
    f.write(keras_model_json)

# 读取Keras模型的结构
with open('model.json', 'r') as f:
    keras_model_json = f.read()

# 将Keras模型的结构转换为K210的.kmodel模型
k210_model = model_from_json(keras_model_json)
k210_model.load_weights('weights.h5')

# 保存K210的.kmodel模型
k210_model.save('model.kmodel')

注意：此代码只适用于具有以下操作的.tflite模型：CONV_2D，DEPTHWISE_CONV_2D，AVERAGE_POOL_2D，MAX_POOL_2D，ADD，RELU，RESHAPE，FULLY_CONNECTED和SOFTMAX。如果您的模型包含其他操作，则需要对代码进行修改。