将训练好的tensorflow模型移植到android应用中

具体步骤如下：

1.  TFLiteConverter保存模型

      修改网络模型代码，将模型通过TFLiteConverter转化成为 TensorFlow Lite FlatBuffer即为.tflite的备份文件。参考官网说明https://tensorflow.google.cn/lite/convert/python_api

　　这里我选择的模型是tensorflow tutorial里面的mnist代码，原因是比较简单，方便实验。具体路径models-master/tutorials/image/mnist/

　　对于输入、输出，我做了一下修改，简化为一张图片的预测。

     输入节点：

  eval_data1 = tf.placeholder(
      data_type(),
      shape=(1, IMAGE_SIZE, IMAGE_SIZE, 1))

    输出节点：

  eval_prediction1 = tf.nn.softmax(model(eval_data1))

    用converter保存模型：

        if FLAGS.save_tflite:
            # save tflite
            converter = tf.lite.TFLiteConverter.from_session(sess, [eval_data1], [eval_prediction1])
            tflite_model = converter.convert()
            open("converted_model.tflite", "wb").write(tflite_model)
            print('===========save tflite over =============')

    完整代码：

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""Simple, end-to-end, LeNet-5-like convolutional MNIST model example.

This should achieve a test error of 0.7%. Please keep this model as simple and
linear as possible, it is meant as a tutorial for simple convolutional models.
Run with --self_test on the command line to execute a short self-test.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import gzip
import os
import sys
import time

import numpy
from six.moves import urllib
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

# CVDF mirror of http://yann.lecun.com/exdb/mnist/
SOURCE_URL = 'https://storage.googleapis.com/cvdf-datasets/mnist/'
WORK_DIRECTORY = 'data'
IMAGE_SIZE = 28
NUM_CHANNELS = 1
PIXEL_DEPTH = 255
NUM_LABELS = 10
VALIDATION_SIZE = 5000  # Size of the validation set.
SEED = 66478  # Set to None for random seed.
BATCH_SIZE = 64
NUM_EPOCHS = 10
EVAL_BATCH_SIZE = 64
EVAL_FREQUENCY = 100  # Number of steps between evaluations.


FLAGS = None


def data_type():
  """Return the type of the activations, weights, and placeholder variables."""
  if FLAGS.use_fp16:
    return tf.float16
  else:
    return tf.float32


def maybe_download(filename):
  """Download the data from Yann's website, unless it's already here."""
  if not tf.gfile.Exists(WORK_DIRECTORY):
    tf.gfile.MakeDirs(WORK_DIRECTORY)
  filepath = os.path.join(WORK_DIRECTORY, filename)
  if not tf.gfile.Exists(filepath):
    filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
    with tf.gfile.GFile(filepath) as f:
      size = f.size()
    print('Successfully downloaded', filename, size, 'bytes.')
  return filepath


def extract_data(filename, num_images):
  """Extract the images into a 4D tensor [image index, y, x, channels].

  Values are rescaled from [0, 255] down to [-0.5, 0.5].
  """
  print('Extracting', filename)
  with gzip.open(filename) as bytestream:
    bytestream.read(16)
    buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images * NUM_CHANNELS)
    data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32)
    data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
    data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS)
    return data


def extract_labels(filename, num_images):
  """Extract the labels into a vector of int64 label IDs."""
  print('Extracting', filename)
  with gzip.open(filename) as bytestream:
    bytestream.read(8)
    buf = bytestream.read(1 * num_images)
    labels = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.int64)
  return labels


def fake_data(num_images):
  """Generate a fake dataset that matches the dimensions of MNIST."""
  data = numpy.ndarray(
      shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
      dtype=numpy.float32)
  labels = numpy.zeros(shape=(num_images,), dtype=numpy.int64)
  for image in xrange(num_images):
    label = image % 2
    data[image, :, :, 0] = label - 0.5
    labels[image] = label
  return data, labels


def error_rate(predictions, labels):
  """Return the error rate based on dense predictions and sparse labels."""
  return 100.0 - (
      100.0 *
      numpy.sum(numpy.argmax(predictions, 1) == labels) /
      predictions.shape[0])


def main(_):
  if FLAGS.self_test:
    print('Running self-test.')
    train_data, train_labels = fake_data(256)
    validation_data, validation_labels = fake_data(EVAL_BATCH_SIZE)
    test_data, test_labels = fake_data(EVAL_BATCH_SIZE)
    num_epochs = 1
  else:
    # Get the data.
    train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
    train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
    test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
    test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')

    # Extract it into numpy arrays.
    train_data = extract_data(train_data_filename, 60000)
    train_labels = extract_labels(train_labels_filename, 60000)
    test_data = extract_data(test_data_filename, 10000)
    test_labels = extract_labels(test_labels_filename, 10000)

    # Generate a validation set.
    validation_data = train_data[:VALIDATION_SIZE, ...]
    validation_labels = train_labels[:VALIDATION_SIZE]
    train_data = train_data[VALIDATION_SIZE:, ...]
    train_labels = train_labels[VALIDATION_SIZE:]
    num_epochs = NUM_EPOCHS
  train_size = train_labels.shape[0]

  # This is where training samples and labels are fed to the graph.
  # These placeholder nodes will be fed a batch of training data at each
  # training step using the {feed_dict} argument to the Run() call below.
  train_data_node = tf.placeholder(
      data_type(),
      shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
  train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
  eval_data = tf.placeholder(
      data_type(),
      shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
  eval_data1 = tf.placeholder(
      data_type(),
      shape=(1, IMAGE_SIZE, IMAGE_SIZE, 1))

  # The variables below hold all the trainable weights. They are passed an
  # initial value which will be assigned when we call:
  # {tf.global_variables_initializer().run()}
  conv1_weights = tf.Variable(
      tf.truncated_normal([5, 5, NUM_CHANNELS, 32],  # 5x5 filter, depth 32.
                          stddev=0.1,
                          seed=SEED, dtype=data_type()))
  conv1_biases = tf.Variable(tf.zeros([32], dtype=data_type()))
  conv2_weights = tf.Variable(tf.truncated_normal(
      [5, 5, 32, 64], stddev=0.1,
      seed=SEED, dtype=data_type()))
  conv2_biases = tf.Variable(tf.constant(0.1, shape=[64], dtype=data_type()))
  fc1_weights = tf.Variable(  # fully connected, depth 512.
      tf.truncated_normal([IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512],
                          stddev=0.1,
                          seed=SEED,
                          dtype=data_type()))
  fc1_biases = tf.Variable(tf.constant(0.1, shape=[512], dtype=data_type()))
  fc2_weights = tf.Variable(tf.truncated_normal([512, NUM_LABELS],
                                                stddev=0.1,
                                                seed=SEED,
                                                dtype=data_type()))
  fc2_biases = tf.Variable(tf.constant(
      0.1, shape=[NUM_LABELS], dtype=data_type()))

  # We will replicate the model structure for the training subgraph, as well
  # as the evaluation subgraphs, while sharing the trainable parameters.
  def model(data, train=False):
    """The Model definition."""
    # 2D convolution, with 'SAME' padding (i.e. the output feature map has
    # the same size as the input). Note that {strides} is a 4D array whose
    # shape matches the data layout: [image index, y, x, depth].
    conv = tf.nn.conv2d(data,
                        conv1_weights,
                        strides=[1, 1, 1, 1],
                        padding='SAME')
    # Bias and rectified linear non-linearity.
    relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
    # Max pooling. The kernel size spec {ksize} also follows the layout of
    # the data. Here we have a pooling window of 2, and a stride of 2.
    pool = tf.nn.max_pool(relu,
                          ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1],
                          padding='SAME')
    conv = tf.nn.conv2d(pool,
                        conv2_weights,
                        strides=[1, 1, 1, 1],
                        padding='SAME')
    relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
    pool = tf.nn.max_pool(relu,
                          ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1],
                          padding='SAME')
    # Reshape the feature map cuboid into a 2D matrix to feed it to the
    # fully connected layers.
    pool_shape = pool.get_shape().as_list()
    reshape = tf.reshape(
        pool,
        [pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
    # Fully connected layer. Note that the '+' operation automatically
    # broadcasts the biases.
    hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
    # Add a 50% dropout during training only. Dropout also scales
    # activations such that no rescaling is needed at evaluation time.
    if train:
      hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
    return tf.matmul(hidden, fc2_weights) + fc2_biases

  # Training computation: logits + cross-entropy loss.
  logits = model(train_data_node, True)
  loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
      labels=train_labels_node, logits=logits))

  # L2 regularization for the fully connected parameters.
  regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
                  tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
  # Add the regularization term to the loss.
  loss += 5e-4 * regularizers

  # Optimizer: set up a variable that's incremented once per batch and
  # controls the learning rate decay.
  batch = tf.Variable(0, dtype=data_type())
  # Decay once per epoch, using an exponential schedule starting at 0.01.
  learning_rate = tf.train.exponential_decay(
      0.01,                # Base learning rate.
      batch * BATCH_SIZE,  # Current index into the dataset.
      train_size,          # Decay step.
      0.95,                # Decay rate.
      staircase=True)
  # Use simple momentum for the optimization.
  optimizer = tf.train.MomentumOptimizer(learning_rate,
                                         0.9).minimize(loss,
                                                       global_step=batch)

  # Predictions for the current training minibatch.
  train_prediction = tf.nn.softmax(logits)

  # Predictions for the test and validation, which we'll compute less often.
  eval_prediction = tf.nn.softmax(model(eval_data))

  eval_prediction1 = tf.nn.softmax(model(eval_data1))

  # Small utility function to evaluate a dataset by feeding batches of data to
  # {eval_data} and pulling the results from {eval_predictions}.
  # Saves memory and enables this to run on smaller GPUs.
  def eval_in_batches(data, sess):
    """Get all predictions for a dataset by running it in small batches."""
    size = data.shape[0]
    if size < EVAL_BATCH_SIZE:
      raise ValueError("batch size for evals larger than dataset: %d" % size)
    predictions = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.float32)
    for begin in xrange(0, size, EVAL_BATCH_SIZE):
      end = begin + EVAL_BATCH_SIZE
      if end <= size:
        predictions[begin:end, :] = sess.run(
            eval_prediction,
            feed_dict={eval_data: data[begin:end, ...]})
      else:
        batch_predictions = sess.run(
            eval_prediction,
            feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
        predictions[begin:, :] = batch_predictions[begin - size:, :]
    return predictions

  # Create a local session to run the training.
  start_time = time.time()
  with tf.Session() as sess:
    # Run all the initializers to prepare the trainable parameters.
    tf.global_variables_initializer().run()
    print('Initialized!')

    if FLAGS.use_tflite:
        interpreter = tf.lite.Interpreter(model_path="converted_model.tflite")
        interpreter.allocate_tensors()

        input_details = interpreter.get_input_details()
        output_details = interpreter.get_output_details()
        #eval_data = test_data[0:EVAL_BATCH_SIZE, ...]
        #print(test_data[0, ..., 0])
        print("===========")
        # eval_data = numpy.ndarray(
        #     shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
        #     dtype=numpy.float32)
        # #eval_data = test_data[0, ...]
        # interpreter.set_tensor(input_details[0]['index'], test_data[0:EVAL_BATCH_SIZE, ...])

        my_data = numpy.ndarray(
                shape=(1, IMAGE_SIZE, IMAGE_SIZE, 1),
                dtype=numpy.float32)
        my_data[0, :, :, 0] = test_data[1, :, :, 0]
        interpreter.set_tensor(input_details[0]['index'], my_data)

        interpreter.invoke()
        eval_prediction = interpreter.get_tensor(output_details[0]['index'])
        print(eval_prediction)
        print(numpy.argmax(eval_prediction, 1))
        print(test_labels[0:EVAL_BATCH_SIZE, ...])
    else:
        # #restore checkpoint
        # saver = tf.train.Saver()
        # ckpt = tf.train.get_checkpoint_state("./mnist-model/")
        # print('===================')
        # print(ckpt)
        # print('===================')
        # '''saver.restore(sess, ckpt.all_model_checkpoint_paths[0])
        # print (ckpt.all_model_checkpoint_paths[0])'''
        #
        # saver.restore(sess, ckpt.model_checkpoint_path)
        # print(ckpt.model_checkpoint_path)
        # print('===================')

        # Loop through training steps.
        for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
            # Compute the offset of the current minibatch in the data.
            # Note that we could use better randomization across epochs.
            offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
            batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
            batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
            # This dictionary maps the batch data (as a numpy array) to the
            # node in the graph it should be fed to.
            feed_dict = {train_data_node: batch_data,
                   train_labels_node: batch_labels}
            # Run the optimizer to update weights.
            sess.run(optimizer, feed_dict=feed_dict)
            # print some extra information once reach the evaluation frequency
            if step % EVAL_FREQUENCY == 0:

                # #save checkpoint
                # saver.save(sess, "./mnist-model/model.ckpt", global_step=step)

                # fetch some extra nodes' data
                l, lr, predictions = sess.run([loss, learning_rate, train_prediction], feed_dict=feed_dict)
                elapsed_time = time.time() - start_time
                start_time = time.time()
                print('Step %d (epoch %.2f), %.1f ms' %
                    (step, float(step) * BATCH_SIZE / train_size,
                    1000 * elapsed_time / EVAL_FREQUENCY))
                print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
                print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
                print('Validation error: %.1f%%' % error_rate(
                    eval_in_batches(validation_data, sess), validation_labels))
                sys.stdout.flush()
        # Finally print the result!
        test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
        print('Test error: %.1f%%' % test_error)
        if FLAGS.save_tflite:
            # save tflite, if we want to save with multiple inputs and outputs, the format is like [eval_data, eval_data1], [eval_predictiion, eval_prediction1]. 
            # you must notice that each output must can be worked out by inputs, or there may be segment fault. 
            converter = tf.lite.TFLiteConverter.from_session(sess, [eval_data1], [eval_prediction1])
            tflite_model = converter.convert()
            open("converted_model.tflite", "wb").write(tflite_model)
            print('===========save tflite over =============')
        if FLAGS.self_test:
            print('test_error', test_error)
            assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % (
                test_error,)


if __name__ == '__main__':
  parser = argparse.ArgumentParser()
  parser.add_argument(
      '--use_fp16',
      default=False,
      help='Use half floats instead of full floats if True.',
      action='store_true')
  parser.add_argument(
      '--self_test',
      default=False,
      action='store_true',
      help='True if running a self test.')
  parser.add_argument(
      '--use_tflite',
      default=False,
      action='store_true',
      help='True if running by tflite.')
  parser.add_argument(
      '--save_tflite',
      default=False,
      action='store_true',
      help='True if running by tflite.')

  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

2.  创建Android工程

　　创建一个空的android项目，在main下创建assets文件夹，将上一步保存的converted_model.tflite文件拷贝到assets下。

　　修改app下的build.gradle, 加入tensorflow-lite 依赖。

apply plugin: 'com.android.application'

android {
    compileSdkVersion 29
    buildToolsVersion "29.0.1"
    defaultConfig {
        applicationId "com.example.testtflite"
        minSdkVersion 28
        targetSdkVersion 29
        versionCode 1
        versionName "1.0"
        testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
    }
    buildTypes {
        release {
            minifyEnabled false
            proguardFiles getDefaultProguardFile('proguard-android-optimize.txt'), 'proguard-rules.pro'
        }
    }

    android.applicationVariants.all {
        variant ->
            variant.outputs.all {
                outputFileName = "app_test.apk"
            }
    }
}

dependencies {
    implementation fileTree(dir: 'libs', include: ['*.jar'])
    implementation 'androidx.appcompat:appcompat:1.1.0'
    implementation 'androidx.constraintlayout:constraintlayout:1.1.3'
    testImplementation 'junit:junit:4.12'
    androidTestImplementation 'androidx.test:runner:1.2.0'
    androidTestImplementation 'androidx.test.espresso:espresso-core:3.2.0'
    //tflite
    implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly'
}

　　简单修改MainActivity.java

　　

package com.example.testtflite;

import android.content.res.AssetFileDescriptor;
import android.os.Bundle;
import android.util.Log;

import androidx.appcompat.app.AppCompatActivity;

import org.tensorflow.lite.Interpreter;

import java.io.FileInputStream;
import java.io.IOException;
import java.nio.ByteBuffer;
import java.nio.channels.FileChannel;

public class MainActivity extends AppCompatActivity {

    @Override
    protected void onCreate(Bundle savedInstanceState) {
        super.onCreate(savedInstanceState);
        setContentView(R.layout.activity_main);
        Log.i("YY", "begin");
        try {
            AssetFileDescriptor fileDescriptor = getAssets().openFd("converted_model.mp3");
            FileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor());
            FileChannel fileChannel = inputStream.getChannel();
            long startOffset = fileDescriptor.getStartOffset();
            long declaredLength = fileDescriptor.getDeclaredLength();
            ByteBuffer tfliteModel = fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);
            Log.i("YY", "get model");
            Interpreter interpreter = new Interpreter(tfliteModel);
            Log.i("YY", "load model");
            float[][][][] input = new float[1][28][28][1];
            for (int i=0; i<28; i++) {
                for (int j=0; j<28; j++) {

                    if ((i == 7 || i == 8) && (j > 2 && j<14)) {
                        input[0][i][j][0] = 0.5f;
                    }else if (i > 7 && (j == 12 || j== 13)) {
                        input[0][i][j][0] = 0.5f;
                    } else {
                        input[0][i][j][0] = -0.5f;
                    }
                }
            }
            float[][] output = new float[1][10];
            interpreter.run(input, output);
            float out, maxOut=0;
            int max_index = -1;
            for (int i=0; i<10; i++) {
                 out = output[0][i];
                 if(maxOut < out) {
                     maxOut = out;
                     max_index = i;
                 }
//                Log.i("YY", Float.toString(out));

            }
            Log.i("YY", "predict number : " + max_index);

            Log.i("YY", "finish");

        } catch (IOException e) {
            e.printStackTrace();
            Log.e("YY", e.getMessage());
        }

    }
}

　　这中间启动app报了一个错误，java.io.FileNotFoundException: This file can not be opened as a file descriptor; it is probably compressed， 查了一下，意思是文件没有找到，或者文件被压缩了，原来是打包apk的时候，相当于将assets的文件添加到了一个.zip文件里，进行了一些压缩，导致openFd方法找不到此文件的描述。参考http://ponystyle.com/blog/2010/03/26/dealing-with-asset-compression-in-android-apps/

　　修改文件后缀名为mp3, 可以正常运行。