一 模型保存和恢复

1.1 保存为pb文件和从pb文件恢复

  1. ckpt = tf.train.latest_checkpoint(ckpt_dir)
  2.         cnn = TextCNN(
  3.             w2v_model,
  4.             sequence_length=30,
  5.             num_classes=2,
  6.             vocab_size=352218,
  7.             embedding_size=300,
  8.             filter_sizes=[2,3,4],
  9.             num_filters=128,
  10.             l2_reg_lambda=0.0)
  11.         issue = tf.InteractiveSession()
  12.         issue.run(tf.global_variables_initializer())
  13.         saver = tf.train.Saver()
  14.         saver.restore(issue, ckpt)
  15.         constant_graph = graph_util.convert_variables_to_constants(issue, issue.graph_def, ['output/predictions', 'output/scores'])
  16.         constant_graph = graph_util.remove_training_nodes(constant_graph)
  18.         with tf.gfile.GFile('./pb/model.pb', mode='wb') as f:
  19.             f.write(constant_graph.SerializeToString())
  1. graph_def = tf.GraphDef()
  2.     with tf.gfile.FastGFile(pb_file, 'rb') as model_f:
  3.         graph_def.ParseFromString(model_f.read())
  4.         _ = tf.import_graph_def(graph_def, name='')
  5.     input_x = tf.get_default_graph().get_tensor_by_name('input_x:0')
  6.     dropout_prob = tf.get_default_graph().get_tensor_by_name('dropout_keep_prob:0')
  7.     predictions = tf.get_default_graph().get_tensor_by_name('output/predictions:0')
  8.     scores = tf.get_default_graph().get_tensor_by_name('output/scores:0')
  9.     isess = tf.InteractiveSession()

1.2 ckpt的保存和恢复