17

我正在尝试使用我自己的数据集和类对在 imagenet 上预训练的 Inception-resnet v2 模型进行迁移学习。我的原始代码库是对tf.slim我再也找不到的示例的修改，现在我正在尝试使用该tf.estimator.*框架重写相同的代码。

但是，我遇到了一个问题，即从预训练的检查点只加载一些权重，用它们的默认初始化程序初始化剩余的层。

研究这个问题，我发现了这个 GitHub 问题和这个问题，都提到需要tf.train.init_from_checkpoint在我的model_fn. 我试过了，但鉴于两者都没有例子，我想我做错了。

这是我的最小示例：

import sys
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
import tensorflow as tf
import numpy as np

import inception_resnet_v2

NUM_CLASSES = 900
IMAGE_SIZE = 299

def input_fn(mode, num_classes, batch_size=1):
  # some code that loads images, reshapes them to 299x299x3 and batches them
  return tf.constant(np.zeros([batch_size, 299, 299, 3], np.float32)), tf.one_hot(tf.constant(np.zeros([batch_size], np.int32)), NUM_CLASSES)


def model_fn(images, labels, num_classes, mode):
  with tf.contrib.slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope()):
    logits, end_points = inception_resnet_v2.inception_resnet_v2(images,
                                             num_classes, 
                                             is_training=(mode==tf.estimator.ModeKeys.TRAIN))
  predictions = {
      'classes': tf.argmax(input=logits, axis=1),
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)

  exclude = ['InceptionResnetV2/Logits', 'InceptionResnetV2/AuxLogits']
  variables_to_restore = tf.contrib.slim.get_variables_to_restore(exclude=exclude)
  scopes = { os.path.dirname(v.name) for v in variables_to_restore }
  tf.train.init_from_checkpoint('inception_resnet_v2_2016_08_30.ckpt',
                                {s+'/':s+'/' for s in scopes})
  
  tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
  total_loss = tf.losses.get_total_loss()    #obtain the regularization losses as well
  
  # Configure the training op
  if mode == tf.estimator.ModeKeys.TRAIN:
    global_step = tf.train.get_or_create_global_step()
    optimizer = tf.train.AdamOptimizer(learning_rate=0.00002)
    train_op = optimizer.minimize(total_loss, global_step)
  else:
    train_op = None
  
  return tf.estimator.EstimatorSpec(
    mode=mode,
    predictions=predictions,
    loss=total_loss,
    train_op=train_op)

def main(unused_argv):
  # Create the Estimator
  classifier = tf.estimator.Estimator(
      model_fn=lambda features, labels, mode: model_fn(features, labels, NUM_CLASSES, mode),
      model_dir='model/MCVE')

  # Train the model  
  classifier.train(
      input_fn=lambda: input_fn(tf.estimator.ModeKeys.TRAIN, NUM_CLASSES, batch_size=1),
      steps=1000)
    
  # Evaluate the model and print results
  eval_results = classifier.evaluate(
      input_fn=lambda: input_fn(tf.estimator.ModeKeys.EVAL, NUM_CLASSES, batch_size=1))
  print()
  print('Evaluation results:\n    %s' % eval_results)
 
if __name__ == '__main__':
  tf.app.run(main=main, argv=[sys.argv[0]])

Tensorflow 的模型存储库中的模型实现inception_resnet_v2在哪里。

如果我运行这个脚本，我会从 获得一堆信息日志init_from_checkpoint，但是，在会话创建时，它似乎试图Logits从检查点加载权重，但由于形状不兼容而失败。这是完整的追溯：

Traceback (most recent call last):

  File "<ipython-input-6-06fadd69ae8f>", line 1, in <module>
    runfile('C:/Users/1/Desktop/transfer_learning_tutorial-master/MCVE.py', wdir='C:/Users/1/Desktop/transfer_learning_tutorial-master')

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\spyder\utils\site\sitecustomize.py", line 710, in runfile
    execfile(filename, namespace)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\spyder\utils\site\sitecustomize.py", line 101, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "C:/Users/1/Desktop/transfer_learning_tutorial-master/MCVE.py", line 77, in <module>
    tf.app.run(main=main, argv=[sys.argv[0]])

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\platform\app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))

  File "C:/Users/1/Desktop/transfer_learning_tutorial-master/MCVE.py", line 68, in main
    steps=1000)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\estimator\estimator.py", line 302, in train
    loss = self._train_model(input_fn, hooks, saving_listeners)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\estimator\estimator.py", line 780, in _train_model
    log_step_count_steps=self._config.log_step_count_steps) as mon_sess:

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 368, in MonitoredTrainingSession
    stop_grace_period_secs=stop_grace_period_secs)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 673, in __init__
    stop_grace_period_secs=stop_grace_period_secs)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 493, in __init__
    self._sess = _RecoverableSession(self._coordinated_creator)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 851, in __init__
    _WrappedSession.__init__(self, self._create_session())

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 856, in _create_session
    return self._sess_creator.create_session()

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 554, in create_session
    self.tf_sess = self._session_creator.create_session()

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\monitored_session.py", line 428, in create_session
    init_fn=self._scaffold.init_fn)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\training\session_manager.py", line 279, in prepare_session
    sess.run(init_op, feed_dict=init_feed_dict)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\client\session.py", line 889, in run
    run_metadata_ptr)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\client\session.py", line 1120, in _run
    feed_dict_tensor, options, run_metadata)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\client\session.py", line 1317, in _do_run
    options, run_metadata)

  File "C:\Users\1\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow\python\client\session.py", line 1336, in _do_call
    raise type(e)(node_def, op, message)

InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [900] rhs shape= [1001]    [[Node: Assign_1145 = Assign[T=DT_FLOAT,
_class=["loc:@InceptionResnetV2/Logits/Logits/biases"], use_locking=true, validate_shape=true,
_device="/job:localhost/replica:0/task:0/device:CPU:0"](InceptionResnetV2/Logits/Logits/biases, checkpoint_initializer_1145)]]

使用时我做错了什么init_from_checkpoint？我们究竟应该如何在我们的“使用”它model_fn？Logits当我明确告诉它不要时，为什么估计器试图从检查点加载' 权重？

更新：

在评论中的建议之后，我尝试了其他方法来调用tf.train.init_from_checkpoint.

使用{v.name: v.name}

如果按照评论中的建议，我将调用替换为{v.name:v.name for v in variables_to_restore}，我会收到此错误：

ValueError: Assignment map with scope only name InceptionResnetV2/Conv2d_2a_3x3 should map
to scope only InceptionResnetV2/Conv2d_2a_3x3/weights:0. Should be 'scope/': 'other_scope/'.

使用{v.name: v}

相反，如果我尝试使用name:variable映射，则会收到以下错误：

ValueError: Tensor InceptionResnetV2/Conv2d_2a_3x3/weights:0 is not found in
inception_resnet_v2_2016_08_30.ckpt checkpoint
{'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean': [256], 
'InceptionResnetV2/Repeat/block35_9/Branch_0/Conv2d_1x1/BatchNorm/beta': [32], ...

该错误继续列出我认为是检查点中的所有变量名称（或者它可能是范围？）。

更新 (2)

在检查了上面的最新错误后，我看到它InceptionResnetV2/Conv2d_2a_3x3/weights 在检查点变量列表中。问题是:0最后！ 我现在将验证这是否确实解决了问题，如果是这样，我会发布答案。

4

2 回答 2

9

感谢@KathyWu 的评论，我走上了正确的道路并发现了问题。

实际上，我计算的方式scopes将包括InceptionResnetV2/范围，这将触发范围“下”所有变量的加载（即网络中的所有变量）。然而，用正确的字典替换它并非易事。

在可能的范围模式init_from_checkpoint中，我必须使用的一种是'scope_variable_name': variable一种，但没有使用实际的variable.name属性。

variable.name看起来像'some_scope/variable_name:0'：。这:0不在检查点变量的名称中，因此使用scopes = {v.name:v.name for v in variables_to_restore}将引发“未找到变量”错误。

使其工作的技巧是从名称中剥离张量索引：

tf.train.init_from_checkpoint('inception_resnet_v2_2016_08_30.ckpt', 
                              {v.name.split(':')[0]: v for v in variables_to_restore})
于 2017-12-19T09:12:33.900 回答
1

我发现{s+'/':s+'/' for s in scopes}没有用，只是因为variables_to_restore包含类似"global_step", 所以范围包括可以包括所有内容的全局范围。你需要打印variables_to_restore，找到"global_step"东西，然后把它放进去"exclude"。

于 2018-06-12T17:02:41.107 回答