導出SaveModel通用模型

• SavedModel格式

  SavedModel是目前官方推薦的模型保存的格式（SessionBundle自Tensorflow 1.0以后不再推薦使用），目錄結構如下。

  assets/assets.extra/variables/ variables.data-xxxxx-of-xxxxx variables.indexsaved_model.pb

  目錄中各個子目錄和文件的含義請參見TensorFlow SavedModel官方文檔或介紹。

• 導出SavedModel

  代碼片段：

  class Softmax(object):
      def __init__(self):
          self.weights_ = tf.Variable(tf.zeros([FLAGS.image_size, FLAGS.num_classes]),
                  name='weights')
          self.biases_ = tf.Variable(tf.zeros([FLAGS.num_classes]),
                  name='biases')
  
      def signature_def(self):
          images = tf.placeholder(tf.uint8, [None, FLAGS.image_size],
              name='input')
          normalized_images = tf.scalar_mul(1.0 / FLAGS.image_depth,
              tf.to_float(images))
          scores = self.scores(normalized_images)
          tensor_info_x = tf.saved_model.utils.build_tensor_info(images)
          tensor_info_y = tf.saved_model.utils.build_tensor_info(scores)
          return tf.saved_model.signature_def_utils.build_signature_def(
                  inputs={'images': tensor_info_x},
                  outputs={'scores': tensor_info_y},
                  method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
      def savedmodel(self, sess, signature, path):
          export_dir = os.path.join(path, str(FLAGS.model_version))
          builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
          builder.add_meta_graph_and_variables(
              sess, [tf.saved_model.tag_constants.SERVING],
              signature_def_map={
                  'predict_images':
                      signature,
              },
              clear_devices=True)
          builder.save()
  
  model = Softmax()
  signature = model.signature_def()
  
  model.savedmodel(sess, signature, mnist.export_path())

  代碼說明：

  • Softmax類封裝了機器學習模型，其中weights和biases是其最主要的模型參數。

  • signature_def方法描述了預測時，如何從一個placeholder經過數據標準化和前向計算得到輸出的邏輯，并分別作為輸入和輸出構建出一個SignatureDef。

  導出SavedModel至OSS：

  訓練并導出模型的命令如下。

  PAI -name tensorflow
      -Dscript="file://path/to/mnist_savedmodel_oss.py"
      -Dbuckets="oss://mnistdataset/?host=oss-test.aliyun-inc.com&role_arn=acs:ram::127488******:role/odps"
      -DcheckpointDir="oss://mnistdataset/?host=oss-test.aliyun-inc.com&role_arn=acs:ram::127488*********:role/odps";

保存和恢復檢查點

• Checkpoint存儲

  非交互式TensorFlow存儲模型的示例程序如下。

  # -*- coding: utf-8 -*-
  # usage
  # pai -name tensorflow -DcheckpointDir="oss://tftest/examples/?host=oss-test.aliyun-inc.com&role_arn=acs:ram::****:role/odps" -Dscript="file:///path/to/save_model.py";
  import tensorflow as tf
  import json
  import os
  tf.app.flags.DEFINE_string("checkpointDir", "", "oss info")
  FLAGS = tf.app.flags.FLAGS
  print("checkpoint dir:" + FLAGS.checkpointDir)
  # 定義變量
  counter = tf.Variable(1, name="counter")
  one = tf.constant(2)
  sum = tf.add(counter, one)
  new_counter = tf.assign(counter, sum)
  saver = tf.train.Saver()
  init_op = tf.global_variables_initializer()
  with tf.Session() as sess:
      sess.run(init_op)
      coord = tf.train.Coordinator()
      threads = tf.train.start_queue_runners(coord=coord)
      ret = sess.run(new_counter)
      print("Counter:%d" % ret)
      ckp_path = os.path.join(FLAGS.checkpointDir, "model.ckpt")
      save_path = saver.save(sess, ckp_path)
      print("Model saved in file: %s" % save_path)
      coord.request_stop()
      coord.join(threads)

  tf.app.flags.DEFINE_string()和tf.app.flags.FLAGS可以獲取PAI命令中的checkpointDir參數，checkpointDir指定了模型將要存儲到OSS上。

  以下代碼完成了new_counter的計算，并將名稱為counter的變量存儲到模型中（值為3），save_path = saver.save(sess, ckp_path)將模型寫到OSS指定路徑。

  ret = sess.run(new_counter)
  print("Counter:%d" % ret)
  ckp_path = os.path.join(FLAGS.checkpointDir, "model.ckpt")
  save_path = saver.save(sess, ckp_path)
  print("Model saved in file: %s" % save_path)

• Checkpoint恢復

  TensorFlow的Saver類也可以用于模型的恢復，TensorFlow恢復模型的示例如下。

  # -*- coding: utf-8 -*-
  # usage
  # pai -name tensorflow -Dbuckets="oss://tftest/examples/?host=oss-test.aliyun-inc.com&role_arn=acs:ram::***:role/odps" -Dscript="file:///path/to/restore_model.py";
  import tensorflow as tf
  import json
  import os
  tf.app.flags.DEFINE_string("buckets", "", "oss info")
  FLAGS = tf.app.flags.FLAGS
  print("buckets:" + FLAGS.buckets)
  # 定義變量
  counter = tf.Variable(1, name="counter")
  saver = tf.train.Saver()
  init_op = tf.global_variables_initializer()
  with tf.Session() as sess:
      sess.run(init_op)
      coord = tf.train.Coordinator()
      threads = tf.train.start_queue_runners(coord=coord)
      ret = sess.run(counter)
      print("Before restore counter:%d" % ret)
      print("Model restore from file")
      ckp_path = os.path.join(FLAGS.buckets, "model.ckpt")
      saver.restore(sess, ckp_path)
      ret = sess.run(counter)
      print("After restore counter:%d" % ret)
      coord.request_stop()
      coord.join(threads)

  tf.app.flags.DEFINE_string()和tf.app.flags.FLAGS可以獲取PAI命令中的buckets參數，buckets指定了模型將要從OSS上恢復模型。

  以下代碼中，首先定義了名稱counter的變量，初始值為1。調用saver.restore(sess, ckp_path)，根據給定的OSS路徑恢復已存儲的模型，最后執行ret = sess.run(counter)得到恢復后的變量的值也是3。

  ret = sess.run(counter)
  print("Before restore counter:%d" % ret)
  print("Model restore from file")
  ckp_path = os.path.join(FLAGS.buckets, "model.ckpt")
  saver.restore(sess, ckp_path)
  ret = sess.run(counter)
  print("After restore counter:%d" % ret)

TF模型部署到EAS