Question

我有一个已保存的图表定义，该定义是使用tf.train.import_meta_graph导入的。该图包含py_func op，它不可序列化。我可以在不构建图表的情况下为此操作定义和分配python函数吗？

Answer 1

这是可能的，但可能有点脆弱。特别是，pyfunc需要按照它们在原始图中定义的顺序重新定义（以便它们在FuncRegistry中具有相同的标识符。）

一个例子。我们可以定义一个包含py_func：

的图形

import tensorflow as tf

def my_py_func(x):
  return 13. * x + 2.

def train_model():
  with tf.Graph().as_default():
    some_input = tf.constant([[1., 2., 3., 4.],
                              [5., 6., 7., 8.]])
    after_py_func = tf.py_func(my_py_func, [some_input], Tout=tf.float32,
                               name="my_py_func")
    coefficient = tf.get_variable(
        "coefficient",
        shape=[])
    bias = tf.get_variable(
        "bias",
        shape=[])
    loss = tf.reduce_sum((coefficient * some_input + bias - after_py_func) ** 2)
    global_step = tf.contrib.framework.get_or_create_global_step()
    train_op = tf.group(tf.train.AdamOptimizer(0.1).minimize(loss),
                        tf.assign_add(global_step, 1))
    # Make it easy to retreive things we care about when the metagraph is reloaded.
    tf.add_to_collection('useful_ops', bias)
    tf.add_to_collection('useful_ops', coefficient)
    tf.add_to_collection('useful_ops', loss)
    tf.add_to_collection('useful_ops', train_op)
    tf.add_to_collection('useful_ops', global_step)
    tf.add_to_collection('useful_ops', some_input)
    init_op = tf.global_variables_initializer()
    saver = tf.train.Saver()
    with tf.Session() as session:
      session.run(init_op)
      for i in range(5000):
        (_, evaled_loss, evaled_coefficient, evaled_bias,
         evaled_global_step) = session.run(
             [train_op, loss, coefficient, bias, global_step])
        if i % 1000 == 0:
          print(evaled_global_step, evaled_loss, evaled_coefficient,
                evaled_bias)
      saver.save(session, "./trained_pyfunc_model", global_step=global_step)

这会做一些基本的训练（匹配py_func中的线性函数）：

1 37350.4 -0.0934748 0.193026
1001 19.2717 12.3749 5.40368
2001 0.108373 12.9532 2.2548
3001 8.28227e-06 12.9996 2.00222
4001 3.77258e-09 13.0 2.00004

如果我们在新的Python会话中尝试加载元图而不重新定义pyfunc，我们会收到错误：

def load_model():
  with tf.Graph().as_default():
    saver = tf.train.import_meta_graph("./trained_pyfunc_model-5000.meta")
    bias, coefficient, loss, train_op, global_step, some_input = tf.get_collection('useful_ops')
    #after_py_func = tf.py_func(my_py_func, [some_input], Tout=tf.float32,
    #                           name="my_py_func")
    with tf.Session() as session:
      saver.restore(session, "./trained_pyfunc_model-5000")
      (_, evaled_loss, evaled_coefficient, evaled_bias,
       evaled_global_step) = session.run(
           [train_op, loss, coefficient, bias, global_step])
      print("Restored: ", evaled_global_step, evaled_loss, evaled_coefficient, evaled_bias)

UnknownError（参见上面的回溯）：KeyError：＆＃39; pyfunc_0＆＃39;

但是，只要py_funcs以相同的顺序定义并具有相同的实现，我们应该没问题：

def load_model():
  with tf.Graph().as_default():
    saver = tf.train.import_meta_graph("./trained_pyfunc_model-5000.meta")
    bias, coefficient, loss, train_op, global_step, some_input = tf.get_collection('useful_ops')
    after_py_func = tf.py_func(my_py_func, [some_input], Tout=tf.float32,
                               name="my_py_func")
    with tf.Session() as session:
      saver.restore(session, "./trained_pyfunc_model-5000")
      (_, evaled_loss, evaled_coefficient, evaled_bias,
       evaled_global_step) = session.run(
           [train_op, loss, coefficient, bias, global_step])
      print("Restored: ", evaled_global_step, evaled_loss, evaled_coefficient, evaled_bias)

这可以让我们继续训练，或者我们想要对恢复的模型做任何其他事情：

Restored:  5001 1.77897e-09 13.0 2.00003

请注意，有状态的py_funcs将更难处理：TensorFlow没有保存任何可能与它们相关联的Python变量！

在TensorFlow中导入GraphDef后设置py_func op

1 个答案: