deeplabv3+训练的ckpt模型Android端部署

最新推荐文章于 2025-05-23 18:30:00 发布

氧气邮递员001

最新推荐文章于 2025-05-23 18:30:00 发布

阅读量1.5k

点赞数 2

CC 4.0 BY-SA版权

文章标签： tensorflow 深度学习 android

本文链接：https://blog.youkuaiyun.com/Power392413/article/details/112366179

本文详细指导如何使用 DeeplabV3+ 的训练模型进行量化处理，包括修改train.py和export_model.py脚本，然后转换为pb模型，进一步转化为TFLite模型，并在Android端进行部署。步骤涉及模型量化、模型导出和在实际应用中的适配。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

deeplabv3+训练的ckpt模型Android端部署

接上一篇文章用deeplabV3+训练自己的数据集并导出.pb模型进行可视化测试

一.对deeplabv3+训练的ckpt模型做量化

参考官方文档
修改train.py文件

flags.DEFINE_string('train_logdir', "./datasets/VOC/out",
                    'Where the checkpoint and logs are stored.')
flags.DEFINE_float('base_learning_rate', 3e-5,
                   'The base learning rate for model training.')
flags.DEFINE_integer('training_number_of_steps', 5000,
                     'The number of steps used for training')  ###########  迭代次数
flags.DEFINE_string('tf_initial_checkpoint', "/home/dreamdeck/Downloads/Tensorflow/models-master/research/deeplab/datasets/VOC/trainout/model.ckpt",
                   'The initial checkpoint in tensorflow format.') ##上一步训练好的模型文件
flags.DEFINE_integer(
    'quantize_delay_step', 0,
    'Steps to start quantized training. If < 0, will not quantize model.')

运行train.py

二.量化后的.ckpt模型转.pb模型

修改export_model.py

flags.DEFINE_string('checkpoint_path', "./datasets/VOC/out/model.ckpt", 'Checkpoint path')

flags.DEFINE_string('export_path', "./datasets/VOC/out/frozen_inference_graph.pb",
                    'Path to output Tensorflow frozen graph.')

flags.DEFINE_integer('num_classes', 2, 'Number of classes.')

flags.DEFINE_multi_integer('crop_size', [513, 513],
                           'Crop size [height, width].')
flags.DEFINE_integer(
    'quantize_delay_step', 0,
    'Steps to start quantized training. If < 0, will not quantize model.')

运行export_model.py

三. pb模型转tflite

修改convert_to_tflite.py文件

flags.DEFINE_string('quantized_graph_def_path', "./datasets/VOC/out/frozen_inference_graph.pb",
                    'Path to quantized graphdef.')
flags.DEFINE_string('output_tflite_path', "./datasets/VOC/out/frozen_inference_graph.tflite", 'Output TFlite model path.')
flags.DEFINE_string(
    'input_tensor_name', "MobilenetV2/MobilenetV2/input:0",
    'Input tensor to TFlite model. This usually should be the input tensor to '
    'model backbone.'
)
flags.DEFINE_string(
    'output_tensor_name', 'ArgMax:0',
    'Output tensor name of TFlite model. By default we output the raw semantic '
    'label predictions.'
)
flags.DEFINE_string(
    'test_image_path', "./datasets/VOC/JPEGImages/ADE_train_00000004.jpg",
    'Path to an image to test the consistency between input graphdef / '
    'converted tflite model.'
)

运行convert_to_tflite.py

四.tflite模型部署

1.下载tensorflow-lite-examples-android并用android studio软件打开
2.将tflite模型文件copy到tensorflow-lite-examples-android-master/tensorflow-lite-examples-android-master/example-image-segmentation/src/main/assets文件夹中
3.注释掉download.gradle文件中的

apply from:'../buildscripts/network.gradle'

//def targetModelFile = "src/main/assets/deeplabv3_257_mv_gpu.tflite"
//def modelDownloadUrl = "https://tfhub.dev/tensorflow/lite-model/deeplabv3/1/default/1?lite-format=tflite"
//
//// Required to download TF Lite model for Smart Reply.
//task downloadLiteModel(type: DownloadUrlTask) {
//    sourceUrl = modelDownloadUrl
//    target = file(targetModelFile)
//}

//preBuild.dependsOn downloadLiteModel

修改ImageSegmentationModelExecutor文件

segmentationMasks = ByteBuffer.allocateDirect(1 * imageSize * imageSize * 8)

    private const val imageSegmentationModel = "frozen_inference_graph.tflite"   
    private const val imageSize = 513
    const val NUM_CLASSES = 2
    
    val segmentColors = IntArray(NUM_CLASSES)
    val labelsArrays = arrayOf(
      "background", "sky"
    )

  private fun extraMaskAndItemsFromByteBuffer(buffer: ByteBuffer,
                                              width: Int,
                                              height: Int,
                                              colors: IntArray): Pair<Bitmap, Set<String>> {
    val start = System.currentTimeMillis()

    val pixels = width * height
    val data = IntArray(pixels)
    val itemsFound = HashSet<String>()
    val cacheInMemory = LongArray(width * height)

    buffer.rewind()
    buffer.asLongBuffer().get(cacheInMemory)

    for (y in 0 until height) {
      for (x in 0 until width) {
        var itemIndex = 0

        var c = cacheInMemory[y * width + x].toInt()
        data[y * width + x] = colors[c]
        itemIndex = c


        itemsFound.add(labelsArrays[itemIndex])
      }
    }

    val end = System.currentTimeMillis()

    Logger.debug("convert $pixels pixels' buffer in ${end - start} ms")

    val maskBitmap = com.dailystudio.devbricksx.utils.ImageUtils.intArrayToBitmap(
      data, imageSize, imageSize)

    return Pair(maskBitmap, itemsFound)
  }

修改local.properties文件为自己对应的sdk和ndk

ndk.dir=/home/dreamdeck/Android/Sdk/ndk/22.0.7026061
sdk.dir=/home/dreamdeck/Android/Sdk

修改ImageSegmentationCameraFragment文件

const val MODEL_IMAGE_SIZE = 513

修改TFLiteImageHelper.kt文件

package com.dailystudio.tflite.example.common.image

import android.graphics.Bitmap
import android.util.Log
import java.nio.ByteBuffer
import java.nio.ByteOrder

object TFLiteImageHelper {

    fun bitmapToByteBuffer(
        bitmapIn: Bitmap,
        width: Int,
        height: Int,
        mean: Float = 0.0f,
        std: Float = 255.0f
    ): ByteBuffer {
        val inputImage = ByteBuffer.allocateDirect(1 * width * height * 3)
        inputImage.order(ByteOrder.nativeOrder())
        inputImage.rewind()

        val intValues = IntArray(width * height)
        bitmapIn.getPixels(intValues, 0, width, 0, 0, width, height)
        var pixel = 0
        for (y in 0 until height) {
            for (x in 0 until width) {
                val value = intValues[pixel++]

                // Normalize channel values to [-1.0, 1.0]. This requirement varies by
                // model. For example, some models might require values to be normalized
                // to the range [0.0, 1.0] instead.
//                val a = value shr 16 and 0xFF
//                val b = value shr 8 and 0xFF
//                val c = value and 0xFF
//                inputImage.putFloat(((value shr 16 and 0xFF) - mean) / std)
//                inputImage.putFloat(((value shr 8 and 0xFF) - mean) / std)
//                inputImage.putFloat(((value and 0xFF) - mean) / std)
//                Log.d("TFLiteImageHelper", "y= $y, x= $x, a= $a, b= $b, c= $c")
                inputImage.put(((value shr 16 and 0xFF).toByte()))
                inputImage.put(((value shr 8 and 0xFF).toByte()))
                inputImage.put(((value and 0xFF).toByte()))
            }
        }

        inputImage.rewind()
        return inputImage
    }

}